Low Data Complexity Power/EM 2
 Low Data Complexity Power/EM 2
 The New York Times, Take 2
 Power Analysis Attacks
 Simple Power Analysis
 Other Types of Power Attacks
 Tuning the power model
 8bit microcontroller
 Power Model for Microcontrollers
 Simple Power Analysis of RSA
 Simple Power Analysis of AES
 Advanced Encryption Standard
 The Advanced Encryption Standard
 AES Internals
 AES Power Analysis
 Template Attacks
 Related work
We want to see what is power analysis all about and see a very simple power analysis attack that we can actually perform. Then we will dive into AES and its implementation.
All the things that we will see today are very optimistic and they assume that we have really good measurements, very good understanding and very good luck, but it’s actually quite a little bit harder in practice.
The New York Times, Take 2
The hero for this article is Mister Kocher. He discovered something that can attack smart cards with what’s called power analysis. Probably it was discovered before, but he has discovered it academically.
What did he do? He went to a conference and took people’s smart cards and found out their secret private keys. We saw last week that if we have to sign something and we have the private key you can sign a thing that’s not supposed to be signed and steal money, this was a big deal.
The 1995 timing attack went to the front page of the New York Times and there was a lot of discussion about the discovery. The 1998 result about power analysis only goes to the bottom of page two in one of the supplements and actually nobody has discussed about it a lot.
Why didn’t he get a lot of attention for this power analysis attack? Could be because timing is one thing and power analysis is a little more complicated but in the bottom line we can see all the secrets. To do a timing attack you need a stopwatch, while to power analysis attack you need more sophisticated equipment.
Power Analysis Attacks
What we need to make timing attack? We just need to be able to send request and gets responses to measured how much it took, while the target can be at the other side of the internet (can Amazon Cloud very far away).
What we need to make Power analysis attack? We need to be physically. How do we connect to the device? We need to cut the power supply and connect to it diractly to measure the power consumption. This is a very invasive attack, we need to be very very close. in 1995 let’s assume that it’s true. so if we go to the system architect:” listen there is an a power analysis attack that’s can completely compromised the device. Attacker needs to come to the device end cuts the power supply….” by the time you already lost the system architect because this is not practical. The threat model has to make sense.
nevertheless, the threat model does make sense in a lot of scenarios and we discovered that thing that we was not supposed existing these power analysis and side Channel attack. today we can attack this device and tomorrow we can attack another device.
To learn about power analysis attack you can go and search in the library the “Power Analysis Attacks” book.“Cryptanalytic attacks that allow extraction of secret information from cryptographic devices by exploding their power consumption characteristics”let’s see what can we discover from this definition. First, What we are attacking here? Cryptographic devices. what we are not attacking? cryptographic algorithms. If I told you that I was able to break RSA on my phone by doing power analysis did I break RSA? No we didn’t break RSA, we break the implementation of RSA. What else we are not attacking? The user is not under attack, we are not doing any key logging or human engineering attack or social engineering we are only attacking the device. What’s the cryptographic device? Some kind of crypto, signing or encrypting. What a secret information we wants to extract? Probably we extract the key. And what’s nice about the key? That he has lots of bits. At the start you don’t know anything about the key, If you get half of the key you halfway, if you get the whole key You win. If I want to make the device more secure what do I need to do to key? we make a longer key.
This is very easy to say, but if I’m talking instead of secret information we cracked from the device, like medical information, it’s a little harder to understand what we can do with half of it. How do we make the medical information more secure? What it’s actually means?
Let’s take credit cards, the companies say that you are more secure with credit card. Why do we think that we have a privacy with credit cards? We are giving the credit card number and our name… Looks like we’re giving everything, the saying that we are actually have privacy, a lot of people have in the US have the same name, but you also get a zip code. Every time you pay with the credit card it generate a new credit card number automatically and doesn’t have any digits on him. It is very hard to talk about privacy if it’s not a cryptographic. What’s left from the definition? we are exploiting the power conceptions characteristic.
What are we not exploiting? We’re not doing math and we are not doing something that you can do by algorithms. It’s important to know that exploiting the power at conception characteristics does not have to be actually measuring the power. We saw last week that we can actually measure the power consumption with other methods. When Kotcher wrote his paper he announced three types of power analysis. One of them called Simple power analysis, the other one called differential power analysis. Let’s see the simple power analysis today.
Simple Power Analysis
The simple power analysis means that I’m going to take the measurement of the device, making one measurement or two. With that we are going to get the key from those measurement. What’s nice about this attack that it’s very reasonable attack model. We need to get the power consumption trace ( this is a vector overtime of the power consumption of the device), and we need only one or two traces. This is actually durable in a lot of scenarios even if I have the device for a little time or even if someone is looking at me. We will see the setup that can be used for simple analysis.
When you go to a store in Europe, you can’t give the credit card for the cashier, they give you this terminale and then you need take your credit card insert it and put the pin number. And what is going over here, there is something like a cryptographic computer between your card and terminal. Let’s assume that we want to attack the card, and it is in my possession. This model is very permissive to me and I can do whatever I want, I can do a lot of transactions I can do radiat, I can twist it and even melt it. so attacking the card is very easy.
But if I don’t wants to attack the card? I want to attack the terminal using power analysis. Maybe the terminal has an SSL private key which is used to connect to the Central Center, and this can make a lot of damage and we can be very rich.
We want to do a power analysis on the terminal. This is a nice setup, this is something that looks like a credit card, but it has are wired connected to this fpga and connect to a computer. I will be in a very cold country and I will take this card out from the jacket to my palm. Insert this chip into the terminal, while doing a power analysis on a this terminal. The idea is that I can do it about 1 or two times, but not making it for a thousand times or melt it. I can attack maybe this terminal or a gas station. The fact that we don’t need a lot of traces is actually good.
What are the disadvantages? The problem here is that when I look at this power trace I need to be very very well prepared, and understand what’s really going on in the power trace. Because we get a vector with a hundred thousand points and we need to understand where is the encryption starting? Where it is ending? What it means that there is a lack of power here a little power there? We need to have a very good understanding of the device processes.
Not only that we need also to have a very good measurements, because I only have one or two measurements. They’re a lot of external Influence on the device, there can be noise, may be related to the device may be related to the environment, and if I have only one measurement I am going to get all of this noise at once and cannot do anything to reduce it. Statistically I’am not going to be able to get clean measurements to perform the attack.
Another problem is that I will need to work very hard to find the key. Let see an example: Yossi did a power analysis measurements and he got a Trace and there was a noise, he did his analysis and got the key. Now he want to check if this is the right key. How can we check the key? Try to decrypt and encrypt the private key. Is this the right key? no it’s not :( .. Why? We have only one Trace. Maybe there was noise? Maybe one measurement was wrong? How do we recover from this? We can try a similar key and not there exact key, maybe to change you on bit here or there. This search might be so intensive that we getting basically the same effect like a Brute Force. This makes simple power analysis are not very effective because of all of those reasons.
Other Types of Power Attacks
So in general in power analysis there are two classes:
Low data complexity attacks
where I get line trace or two traces (a very small amount of traces). Then I do a lot of postprocessing and think really really hard and maybe do a reverse engineering before.
High data complexity attacks
I will talk to you about it next week. These attacks require a lot of traces, maybe thousands or Millions traces, a Terabyte of data.
Tuning the power model
The first thing that you need to do for a simple power analysis attack is to understand what device you are attacking. We attack two general types of devices with simple power analysis, first is a microcontroller or a CPU and the other thing is ASIC.
Microcontroller is basically a regular computer, it gets commands and runs it one after another, if you want to write a new software for this computer you can write it in C or Java, they’re cheap and commonly used.
This is a Bitcoin wallet. The Bitcoin is basically numbers and if someone steals these numbers he steals your money, so you don’t want these numbers to be on your computers. These do all the calculation when you connect to the network, we want it to be very secure because if someone steal the secret key he can take all of your money.
#####
Inside this device it has an orange Square and this is the microcontroller. Her some storage and you can write some code for it.
The other kind of a device is called ASIC, this device is manufactured only for a specific purpose. Let’s say I want to have a sprinkler that starting the morning and ends in the evening, I will use a programming language that called HDL, once I compiled these software I am not getting an executable program that I can run on a device. These chips can only do one thing, I can’t programming them and I can’t upgrade them, if there is a bug I am in a big problem, but they have only one purpose. The power consumption of this devices are going to be smaller and sometimes they even be faster. Microcontrollers have programs that runs one line after another, an ASIC can do a parallel operation. ASIC is very cheap to manufacture, but there is a big wrap up before you produced this.
In this course we are going to talk only about microcontrollers, because we are going to see them more often than ASICs. But you will know enough to open the book attacking ASIC.
So what is the line between microcontroller and an ASIC? On one of the student table there is a chip that he will show us, this is a FPGA field programmable this chip. If I want to identify a particular face I can program this ASIC to do patterning for this face and doing it very very cheaply. I can do also audio compression maybe I don’t know what exactly I want to do but I know that I need to do some kind of compression while I don’t know exactly the algorithm by using this ASIC. So you will find an ASIC if you have a piece of equipment 10000 pieces, maybe a router ,oscilloscope, submarine and things like that. Best to make sure what is an Arduino? Microcontroller.
8bit microcontroller
This is an 8bit microcontroller from the 80s, you see the center of this microcontroller , this pair of trousers the red that named ALU, this is where the magic happens, this piece of silicone can actually do logic like multiply, add, shift or compare.
The entire process of life is to get a line of code which represent an instruction, he has to understand what this instruction is trying to do. It can be multiply, read or write. And then you get the operator you need to do from the memory and fed it to the ALU. Then we’ll get the next instruction and we’ll do it again and again. what’s important for me to show you that there is two long in lines from the top and the bottom they are called the buses. the top called the data bus and the bottom call address bus. What is data bus mean? Any sort of data you need to computation has to fetch from this bus. If something come from the memory it going from this bus. If I want to write to external input output, it also going on this bus. The width of this bus is 8 Bits and it means all of the operation that this microcontroller do our eight bits.
On the bottom there is another big bus which is the address bus. If I want to write to memory, I’m going to put their address in the address bus and then I’m going to put the data in the data bus. I will set the control bus to write, and then the memory which is somewhere else is it going to rise to this address. If I want to do a read, I will put the address I want to read in the address bus, and read in the control. What I will do in the data bus? I don’t want to put something in there because I want to read, this is actually important and I will elaborate about it more later.
Power Model for Microcontrollers
What’s interesting about microcontroller that there are in many cases the power model don’t have Hamming distance and actually have Hamming Weight. What does that mean? That if I think that there are going to be a value going over the bus I don’t need to know what was the previous value on the bus, because it’s going to be exact humming weight for this value.
When you have the best that is shared with several components the idea is that all of the components that are not using the bus are going to set how to put them, so all are ones, So if I want for example to read advice from memory the CPU is going to set everything to one and then he is going to extract the memory from the address, then the memory is going to lower all the relevant bits until what sitting on the memory bus and the data bus what I wanted.
Let’s say I want the memory 4, at first I’m going to see on the data bus 0xFF, then I will see 4 and then going to see agaoxin FF because the memory is finished. what was the powerconsumption here to go from FF to 4, how many beats has to change? 7, and again it goes back to FF.
Simple Power Analysis of RSA
Let’s see the power module this device under test, which does RSA decryption or signing. While it’s doing it we are getting the power measures. Why it is doing an RSA decryption? Because it needs it. We can do it by sending encrypted emails from the phone, so you really have to decrypt the message.
The axis are vector of power measurements, x is time, y are the power consumption,(how did we measure the power consumption? i put a probe, measer the voltege so on.. What is the private key? What is missing? You need to do some reverse engineering first. If this is the only chance to get the key, I need to tell you a little bit more about the device so you will understand what is going on here. This device is microcontroller, it’s doing RSA decryption using right to left binary exponentiation using square and multiply. Is it helping you finding the key? Yes. Let me show you the source code.
Binary exponentiation, it starts from the top most bit, off the secret and privates decryption key, and then for each bit we do Square and multiply. If the bit is zero we do Square, if this bit is one we do square and multiply. This help you now finding the key, let’s look at that Power Trace.
We see two types of things here, this little thing and the big thing. We have some little thing, big thing, little thing, big thing, and then little little little, and then big thing.
All we need to do is to be able tell about were is the square and were is multiply, and then I can read the key, from top to bottom. So telling about square and the multiply to get to the key is the general method. We see square is take a little power and multiply more power.
What is square? square is multiply. So why is the power consumption of square using multiply is different? this is a microcontroller and his bus width is 8 bit. He’s doing a convolution between numbers, so it’s multiply thousands of time in frame. so why this is different? So what is consuming power, the ALU is consumer power, the data bus and the address bus consuming power. In this case the power consumption module of the address bus is hemming distance, because it’s not setting to 1 between accesses it’s always containing what CPU is writing. So did you and multiplication, you are going to see a lot of difference values written into memory to the address bus because this microcontroller have very small component ALU always fetching addresses from memory so ding s*s it’s fetching thousands of thousands of memory, but the address is fetching is very similar to each other, because they are all s. les say s is 1k bits in memory all the above are the same but the bathroom are different.But here we are doing s time m, so it’s fetching s and m, and you doing convolution. so the address bus when it’s doing S and M it’s changing a lot, because he needs to do not only the bottom bits but also the top bits.
Simple Power Analysis of AES
I am going to attack an 8 bit microcontroller. Let’s look on the setup. This microcontroller has a secret key, and it uses an AES encryption. We can ask him to encrypt and decrypt, we send the command in the serial line.
While he’s doing this operation it is going to consume different amount of power and we would like to measure that. How do we going to measure them? We going to connect to the microcontroller power supply. But were do we need to cut and connect? The power is not going straight to the device it is going from this white box (small resistor). How much power is going to go through the resistor? Because he’s small it’s going to take very small power consumption. The voltage drops across this device is going to be measured by this Probe, and I’m going to measure the voltage over time. We know the voltage of the power supplies is 5 volt, so from this we can find out what is the current going through the microcontroller and from the current we can find out what is the power consumption. Do we need the code of the microcontroller? Yes. The only thing I don’t know is what? The secret, but I know everything else.
Advanced Encryption Standard
What is the story about AES the encryption standard.
In Death Valley Days encryption was in military standards it was considered like a weapon you weren’t supposed to have encryption, not by buy, export or sale. but in the seventies the US government what is it might be a good idea to have civilian encryption to protect civilian identities or health. they went to IBM and ask them to write civilian encryption standard. IBM gave them an algorithm called “Lucifer”, which was based on civilian Knowledge from the World War II, the NSA I analyzed it and they say they don’t actually like what you wrote and change it. they changed the key change one of the internal tables and say this was the standard. and the DES actually announced. the changes that NSA made was making it difficult to implement it in software, and to make it Brute Force about using the computing power that has the NSA but to protect it against some kind of attack which was known to the NSA but not known in the Differential cryptography ( not going to teach you).
The time passed and the computer got more and more powerful, there was something called Deep cracking the electronic something, which was able to crack DES at least. I’m not sure if they built it. This was too weak, so introduce something called triple DES, it was twice as secure. Triple DES only used two keys, but we’re not going to study in this course. This wasn’t so very efficient and it was very slow. The US National standard unit, we want to create a new Cipher which was at least as secure as Triple DES but much more efficient. Efficient on software and hardware and slow computers. Anybody could submit candidates, the winner of this competition was AES, which was a PhD work Raymond and Diamond.
It has some very good properties that we are going to talk about them.
The Advanced Encryption Standard
How would you say AES is?, it is an operation which take 2 input, a kye and plain text. And his output is ciphertext. How big is the plain text? 128 bit, The input of the key is not always 16 bits, 16, 64,128 bits. Can go to AES 256 key. No one can break the 256 AES key. What if I wants to decrypt? AES can be a reverse you can put the ciphertext the key and you get the plain text?. Can we get the plain text and ciphertext so we can get the key? In theory we can do it but the idea is that you cannot get the key. When you feed the key it has to work a little bit, has to extend the key, create around key, this is done in very secure facility. What if my input is more than 16 bytes? What if I need to encrypt a file? I need to use a protocol and a mode, I can’t just use AES, only works on 16 bytes. If I have a larger data I need to use AES with a particular way. Something called AES mode, the famous one called ECB, and CBC, the fashionable called GCM. Not in this course and we don’t really care, don’t use ECB. What if we want to encrypt just one byte? we need to do padding, there is a trick and we need to do something. Let’s talk about the design of a AES, it was designed to resist all of source of the attacks, the attacks which were known in the days of DES. And all sorts of attacks which were discovered using the competition. AES was billed to be resistant for those attacks, script analysis attack but no power analysis attack. They are basically expose the cipher if you have enough plaintext and ciphertext.
AES won the competition because it was very fast or efficient. You have three optimization goals when you’re building a crypto implementation. You wanted to be fast and to be able to encrypt as many bits. Maybe you want to encrypt all the data in the router? You need to be fast and you want it to be power efficient. You don’t want to change your battery, and to be cheap so using as little transistors as possible. Take at least area in the Silicon using a smaller cheep. These three goals are conflict with each other, but if you want go hardcore which AES is very very fast, if you want AES be larger. Particular the AES submission the realtime paper head implementation of a s 8 bit 16 and 32 beats microcontroller which be being used till this day. One thing about a s which is very nice is that if you remember the things that people were very very suspicious about this AES that IBM present and a bunch of tables which values that IBM didn’t explain and then the NSA came back sage no so use these different values and they also didn’t explain why. I know that the NSA did something good and what’s nice about AES that he use a single lookup table for all Xbox and this Lucas table is actually derived from mathematical relationship some kind of a multiplication are over algebra field so it’s not so hard there. The design is so simple that you not be able to crypto even if you try.
AES Internals
AES is an iterated cipher Which means eats has a very bASIC algorithm call drafts and he does them all over and over again, AES has 10 Rounds, if you wants to do it more complicated you as more and more rounds. How does as operated, you begin with 16 bytes and put them in a cyber that’s called stage register and then you’re on the round operations on these state bytes, every time you operate operations the plaintext gets mixed with the key and every time you do it it gets more and more. When you finish these 10 Rounds senior stage register you have the ciphertext.
If you want to do it in a reverse you put the ciphertext and the stage register and run the operation run after another and you get the plain text.
Each round consists of 4 basic operations, sub bytes, shift rows, mix columns, and add around key.
Every operation was chosen by the creator Rijndael to achieve a different objective. One of the objectives was to confusion, to make the aisle to put not linear a dependent on the input, and I don’t think they wanted to do what is diffusion so all the output will depend on the input. We are mixing the plaintext with the key with each one of the operation.
SubBytes
First one of the AES is sub bytes. Let’s talk about this while thinking about attacks.
How do you implement on 8bit microcontroller? The state is stored in memory so I have 16 bytes of state, you read the first byte of state and the Sbox. Is a table that stored in memory and the size of the table? 2**8.
So I have this stage registered which is 16 bytes, and the sub box table 256 bytes. A full loop and I took the first byte, first you read it and then I needed to read from the table who is the address that I just read and then I get the value the Sbox, (we know inside the microcontroller) who does right component units the value of the states and the value of the Xbox table, no XOR them, no store that value in the stage register. Bridge from the state go to the table, reading from the sub byte table and xor, and write.This operation is very very leaky.
ShiftRows
Then I need to do shift rows. The first row you don’t need to do anything, the other rows you need to shift them, how do you shift with 8bit microcontroller? using a temporary value, I read the state into the temporal value and I’m right it into here. this is also a leaky operation, because I’m leaking all the bytes in the state, well digging the Hemming Weights of the byte. another way to implement it, just by imagination, it doesn’t change the data so there are actually operations that don’t do shift row.
MixColumns
Next operation mix columns, is the Matrix a complication, perform over algebra algorithm, it’s takes as input 32 bits columns and his output is 32 bytes value what are all of the bytes in the output depend of input. How do you do it on a microcontroller? One of the reasons that rhino run the competition that it was very efficient way was doing mix column 8bit microcontroller, using 13 operations which are shifts and exor. This is the most leaky part of AES this mixed columns.
AddRoundKey
Then do a add round key, just xor, read xor write. This doesn’t leak so much because it is inside the ALU, but the reading and writing is the leaking here. Each round of AES is 84 leaking actions. You take the key and you use the same you used to do in creation but you don’t have the plane text yet so you use sub bytes or shift, then you end with the round, and this key expansion is very sensitive for power analysis, you can really attack as very efficiently. We can assume that this expansion is very secure.
AES Power Analysis
I told you about AES and what is our motivation and what is the structure, and I want to spend the time that has left to show you a little about internal of AES and very very very briefly talk about the reaction of doing simple power analysis of AES.
% Make sure the matlab AES scripts are in the path
addpath('matlab_aes_scripts');
%%
% Create two 128bit plaintexts (exactly 16 byte)
plaintext_1 = uint8('Attack at 12:56!');
plaintext_2 = uint8('Attack at 12:57!');
% how many bits are different between the two?
disp(hamming_weight(bitxor(plaintext_1, plaintext_2)));
%%
% Create a key
key = uint8('1234512345123456');
ENCRYPT = 1;
DECRYPT = 0;
%%
% Encrypt the two plaintexts
ciphertext_1 = aes_crypt_8bit(plaintext_1, key, ENCRYPT);
ciphertext_2 = aes_crypt_8bit(plaintext_2, key, ENCRYPT);
% even though the plaintexts were very similar...
disp([plaintext_1;plaintext_2]);
% ... the ciphertexts are very different
disp([ciphertext_1;ciphertext_2]);
%%
% how many bits are different between the two?
disp(hamming_weight(bitxor(ciphertext_1, ciphertext_2)));
%%
% Decrypt the two ciphertexts
decrypted_1 = aes_crypt_8bit(ciphertext_1, key, DECRYPT);
decrypted_2 = aes_crypt_8bit(ciphertext_2, key, DECRYPT);
% Did we get the plaintext again?
disp([plaintext_1;plaintext_2]);
disp([decrypted_1;decrypted_2]);
%%
% Look at the internals of AES now
[ciphertext_1, leak_1] = aes_crypt_8bit_and_leak(plaintext_1, key, ENCRYPT);
[ciphertext_2, leak_2] = aes_crypt_8bit_and_leak(plaintext_2, key, ENCRYPT);
% Show an image showing the two leaks side by size
subplot(1,3,1)
image(squeeze(leak_1));colormap(hsv(256));
subplot(1,3,3)
image(squeeze(leak_2));colormap(hsv(256));
figure(gcf)
%%
% Show the difference in the middle
subplot(1,3,2)
image(squeeze(bitxor(leak_1,leak_2)));colormap(hsv(256));
figure(gcf)
%%
% plot the HW of the difference
subplot(1,1,1)
bar(hamming_weight(bitxor(leak_1,leak_2)));
What you see here is Matlab, I wrote this lab environment to do AES implementations, There is code that dose AES, and there i code that simulates the power leakages of AES as it was implemented on 8bit microcontroller, all of the operations are 8 bits operations and each time operation happens I’m going to leak it’s Hamming Weight .Let’s see what’s going on.
The first thing I’m going to do is to load some libraries and you can find them in the middle, now going to create 216 bytes plain text. I am going to take these string, to make it binary string Unit8, and I’m going to measure the Hamming distance between these two strings. What is the time distance between these two strings? One, the difference is 6 become 7. 110  111.
How do we do it, I have function called bitxor, and this is a vector of size 16,and then waiting Hamming weight. How many possibles Hamming weight are they for 8 bits value? What can the Hamming weight to be? zero,1,2 … 8.So that Hamming weight here is going to be one. now I’m going to set up a AES I am going to choose a key.And now I’m going to encrypt AES, and then I’m going to use my to plain text and the key. The Hamming distance between the two was one. What will be the Hamming distance with the cipher text? Will it be one?no! what’s possible values it can be, between 0 to 128, because the output. Would you like me to be 128? NO, I would like it to be 64. Why do I don’t need it to be 100 or 2 or 3. Because that means that I don’t have a very important property crypto assistance in the Avalanche property means that each bit in each bit affect all of the input. So if I change one bit in the input and I get only one change in the output it means that I don’t have the ability point. But if I change one bit in the input and I get 100 bit change in the output what it is mean? It’s means that I don’t have the average property and just to make things fun I’m flipping all the bits. What I want the Hamming distance to be is around 64, It’s that exactly half of the bits changing. So you can see I’m doing the Hamming and measuring and show the ciphertext, at the end I’m doing the Hamming weight.
These two strings are very very similar until the end, the 6 and 7 change. But if you see the ciphertext you can see a lot of difference. And this humming weight is 52. sometimes he’s more sometimes is less than 64 but this is okay. Now let’s decrypt, How do I decrypt? I take my AES 8bit function and I give it decrypt. The ciphertext and the key. and see if the plain text into ciphertext are the same. We can see that there are the same and so far so good. Now I want to show you the internal structure of AES. to do that I have a function that’s called 8bit and leak. Let’s open the function. Here is the function. Let’s go over the structure
function [result state rkeys mixcolumn_leak]=
aes_crypt_8bit_and_leak(input_data, secret_key, encrypt)
% performs AES128 encryptions or decryptions like an 8bit uC would do them
% and leaks internal state
%
% DESCRIPTION:
%
% [result state rkeys mixcolumn_leak] = aes_crypt(input_data, secret_key, encrypt)
%
% This function performs an AES128 encryption or decryption of the input
% data with the given secret key.
%
% PARAMETERS:
%
%  input_data:
% A matrix of bytes, where each line consists of a 16 bytes (128 bit)
% data input value of the AES128 en/decryption.
%  secret_key:
% A vector of 16 bytes that represents the secret key.
%  encrypt:
% Paramter indicating whether an encryption or a decryption is performed
% (1=encryption, 0=decryption).
%
% RETURNVALUES:
%
%  result:
% A matrix of bytes of the same size as the byte matrix 'input_data'.
% Each line of this matrix consists of 16 bytes that represent the
% 128bit output of an AES128 en/decryption of the corresponding line of
% 'input_data'.
%  state:
% A matrix of byte of size 'input_data' x 41, containins the state
% progression of the encryption process.
% Legend of the state progression:
% (P= plaintext, C=Ciphertext, K=after AddKey, B=after SubBytes, R=after
% ShiftRows, M=after MixColumns)
% P K BRMK BRMK BRMK BRMK BRMK BRMK BRMK BRMK BRMK BRK(=C)
%  mixcolumn_leak:
% A matrix of size 'input_data' x 9 x 4 x 9 (for encryption), or
% 'input_data' x 9 x 4 x 18 (for decryption),
% where mixcolumn_leak(line, subround, col, :) is the list of
% intermediate valutes generated by the 8bit MC operation on the
% [col] columns of line [line] in the input data during
% subroun [subround]
% EXAMPLE:
%
% result = aes_crypt([1:16; 17:32], 1:16, 1)
% AUTHORS: Stefan Mangard, Mario Kirschbaum, Yossi Oren
%
% CREATION_DATE: 31 July 2001
% LAST_REVISION: 28 October 2008
state = zeros([41 size(input_data)], 'uint8');
rkeys = zeros([10 16], 'uint8');
if (encrypt == 0) % decryption
mixcolumn_leak = zeros([9 4 size(input_data,1) 18]);
else % encryption
mixcolumn_leak = zeros([9 4 size(input_data,1) 9]);
end
for round = 1:10
rkeys(round,:) = aes_round_key(secret_key,round);
end
% expand the keys
if encrypt == 0 %decryption
state(41,:) = input_data;
for i=10:1:1
if i ~= 10
input_data = aes_add_round_key( aes_round_key(secret_key,i), input_data);
state(3 + (i1)*4 + 2,:) = input_data;
[input_data leak] = aes_mix_columns_8bit_and_leak(input_data,0);
mixcolumn_leak(i, :, :, :) = leak;
state(3 + (i1)*4 + 1,:) = input_data;
else
input_data = aes_add_round_key( aes_round_key(secret_key,i), input_data);
state(3 + (i1)*4 + 1,:) = input_data;
end
input_data = aes_shift_rows(input_data,0);
state(3 + (i1)*4,:) = input_data;
input_data = uint8(aes_sbox(input_data,0));
state(3 + (i1)*4  1,:) = input_data;
end
input_data = aes_add_round_key(secret_key, input_data);
state(1,:) = input_data;
else % encryption
state(1,:) = input_data;
input_data = aes_add_round_key(secret_key, input_data);
state(2,:) = input_data;
for i=1:10
input_data = uint8(aes_sbox(input_data,1));
state(3 + (i1)*4,:) = input_data;
input_data = aes_shift_rows(input_data,1);
state(3 + (i1)*4 + 1,:) = input_data;
if i ~= 10
[input_data leak] = aes_mix_columns_8bit_and_leak(input_data,1);
mixcolumn_leak(i, :, :, :) = leak;
state(3 + (i1)*4 + 2,:) = input_data;
input_data = aes_add_round_key( aes_round_key(secret_key,i), input_data);
state(3 + (i1)*4 + 3,:) = input_data;
else
input_data = aes_add_round_key( aes_round_key(secret_key,i), input_data);
state(3 + (i1)*4 + 2,:) = input_data;
end
end
end
result = input_data;
First of all I expand the key, so I do the AES key expansion and I don’t leak the key expansion. This is the assumption. Let’s disregarded that description for a moment and this is the encryption. First of all I took the state and the input data in the state, and then I do a add round key, then do SUB bytes, shift rows, mix column. In the last round I done did you mixed columns, and every time I do this I don’t replace the state and I saved the state. at the end I output the cipher text, but also the output to the state, so you can see the state is always changing. Now let’s see how does it looks, I’m going to run AES twice, and then do some little figure.
The xaxis is the index of the byte in the state. I read the bytes not as a matrix I read it as a row. So there are 16 bites in the state. and the yaxis is the index of the rounds and round 40 is the final round ciphertext. So in between there is the stages. You can see that in the beginning it is very very similar.
If you go down it’s become much different and in the bottom it’s completely different. This is very easy to analyze. And what we are going to do now show the difference between state. What do you see in the middle is the Hemming distance between the right and the left where red is is 0 and blue is 256.
In stage 0 what is the difference between two plain texts? 1 bit, you can see one beat is changed to O. The first round is a add key, we xor the 15 bytes of the key who is 16 bytes of the state, the key is the same in both sides, if the Hamming distance between the two sides was one before I sold the key what is it going to be after? 1. The differential doesn’t change at all, I think a distance of 1 and Ikes or a constant of both sides and The Outpost is still the same. so in the first row I can’t see differences. look wild sheep will do, one by the interstate here, one byte is in the same place and one byte There and one byte here. Now there are 4 bytes different between the states, but this bytes are all of the column.What is the next operation? mixed columns. See what’s mix column did, mixed this change and now all of the bytes of the states are different.We can’t stop AES after 2 rounds, it still can break with crypto analysis, but you can see that this is very nice.
I want to plug the Hamming weight of the distance, The xaxis here he’s there around, and the yaxis is the Hamming distance between the two stay. you can see state with the one then one stays one, after sub byte becomes 4, Shift columns doesn’t change the 4, then mix columns make it grow a little bit. And then you see the distance stays around 64 until the end of the encryption. I want to show you how do we attack AES using simple power analysis.
There is a lot to talk about in very little time. in very very briefly I will give you the ideal way did you it and then I’ll show you how it’s actually done. I have connected a probe to my device, I have measured the power consumption overtime. so now I have a vector of size that lets say a hundred thousand, and I know that AES is there. Now I want to find the key out of my measurements from my trace.
I have a trace, which was recorded in the same device we saw here. I have a vector called 200 of traces of AES encryption, lets plot one of the traces. this is actually only one of the rounds of AES. Here is a trace. This is very very clean trace and I would like it to be in my lab.You can see some of them are high and some of them are low, how do I find the key out of this? let’s start from the beginning then jump to the middle and the end I don’t have time to teach you. the best thing I can hope for is not to get the bytes of the key because the bytes are not function of the bytes, what is the function of? The Hamming weight. So the best hope is the get the Hamming weight, so are you home we will get a vector like the state vector. is this enough to get the key? so you have to believe me that’s yes. how do you do it? You will use algebraic the solver and you will have to read the paper. if you have the vector off all the time in Hamming weight you can get the key. But how do I get the vector Hamming weight from distance? Obviously somewhere in this trace, let’s look at very very leaky operation like sub bytes, or add around key; it read the states you leave the key and xor them together and then you write the state again. Somewhere in this vector there is a peak where we know that exactly in this moment there was a read of the key. The key was read from memory and was travel in the data bus in this exact peak. And how do I find this particular moment in time I will show you in the next lecture.
So I need to write a function that gets an input about 20 points and what is its out puts? the Hamming weight. maybe it will output of Hamming weight. how do I do it? we need to write a decoder that’s as single processing.
Here is a figure showing the move operation they did the same move operation in the microcontroller over and over again this is the average of thousands of measurements where are they moving 0 or 1, until moving 255 ff. Can anybody, you see how Hamming Weight zero have big change, why Hamming weight 0 is more power consumption of Hamming weight 8? Between moves it settings old ones, so moving to zero take more effort from changing.
Here is like a longer trace, let’s assume I hate this data how do I build the decoder. I want the function that can take an array of values and output Hamming weight. Let’s start simple, what if I had just one point? I have an input and function that gets one point and give me the Hamming weight of this point. I can calculate the mean each one, the mean for 0 the mean for 1, for 2. if I get it right what we will do, so do you like a nearest neighbor. if I know that the mean of 4 Hamming weight for is 7 and the mean of Hamming weight five is 8 and I got 7.9, what am I going to chose 4 or 5? 5.nearest neighbor.
This is a nice idea, but I have to give them a little extra dangerous. the problem is here that’s the Hamming weight are not identically distributed. how many possible values of bytes I get? 256. how many bytes are there with Hamming weight zero? 1. how many bytes Hamming weight 8? 1. how many Hamming weight 1? 8. Hamming weight of 1 is 8 times more likely than Hamming weight of 0. So why do we need to do?we need to do Bayes. The idea is are you going to favor the output more likely. I will found out what are the odds that its five, 6?, 7? then look on my decoder and then I’m going to look and give a bonus for more common Hamming weight. The idea is I need to learn How likely bytes based on the trace decoding and then I’m going to go and look at the distribution of this and then multiply The probability I got. this is for one point. I calculated the odds and then I multiply the probability. what if I have more than one point? watch we actually wants to do here he’s actually called multivariate normal distribution. each one of these points, let’s say four points, the dimension, and I have a like a cloud In this distribution space, how do you do this? that is the very very fundamental paper called “template power analysis attack” which explain this.
So if you want to do a simple power analysis on AES the steps you do first of all you profiled the device, you understand where all this stuff is happening, you need to build Template that will help you to recognize different Hamming Weight, then you’re play the same place in the trace and you get guesses for Hamming weight for each States and you hope you get the right guesses, and then you take this Hamming weight and you take a few equation solver we’ll take the Hamming weight and output the key, or something that we don’t know? I don’t know because noise. If we can correlate the Hamming weight to wrong then the equation solver will fail. what can we do in this case? We try again.
Template Attacks
Introduction
Devices performing cryptographic operations can be analyzed by various means. Traditional cryptanalysis looks at the relations between input and output data and the used keys. However, even if the implemented algorithms are secure from a cryptanalysis point of view, sidechannel attacks pose a serious threat. Sidechannel attacks are a subgroup of implementation attacks. Examples thereof are timing attacks, power attacks like DPA or SPA. Traditional DPA style attacks assume the following threat model: The secret key stored in the device is used to perform some cryptographic operations. The attacker monitors these operations using captured sidechannel information like power consumption or electromagnetic emanation. The attack is successful if the used secret key can be reconstructed after a certain number of operations.If the number of operations is limited by the protocol used to initiate these operations, the attacker has an upper bound on the number of operations he can observe. If the operation, which leaks usable sidechannel information, is executed just once, the threat model is different: The attacker has to reconstruct the secret key using a single trace of sidechannel information. Besides protocol limitations, ephemeral keys can be the reason for such a constraint. Techniques like SPA are a general way to tackle this problem. These techniques useeasily distinguishable features of operations like double and add, or add and multiply, to infer keybits. The majority of the available literature deals with these two types of scenarios. If the observed signaltonoise ratio is not high enough, or the implementation is done in a way that ensures the used operations being independent of the key(i. e. no keydependent jumps), SPA style attacks are not possible anymore. The attacker has to think of other ways to get hold of the secret key: One way to do this is to use a similar device and build a model of it. Using this model, an attacker might now be able to recover the secret key.
Algorithm
Template attacks are a powerful type of sidechannel attack. These attacks are a subset of profiling attacks, where an attacker creates a “profile” of a sensitive device and applies this profile to quickly find a victim’s secret key.
Template attacks require more setup than CPA attacks. To perform a template attack, the attacker must have access to another copy of the protected device that they can fully control. Then, they must perform a great deal of preprocessing to create the template  in practice, this may take dozens of thousands of power traces. However, the advantages are that template attacks require a very small number of traces from the victim to complete the attack. With enough preprocessing, the key may be able to be recovered from just a single trace .
There are four steps to a template attack:

Using a copy of the protected device, record a large number of power traces using many different inputs (plaintexts and keys). Ensure that enough traces are recorded to give us information about each subkey value.

Create a template of the device’s operation. This template notes a few “points of interest” in the power traces and a multivariate distribution of the power traces at each point.

the victim device, record a small number of power traces. Use multiple plaintexts. (We have no control over the secret key, which is fixed.)

Apply the template to the attack traces. For each subkey, track which value is most likely to be the correct subkey. Continue until the key has been recovered.
Signals, Noise, and Statistics
Before looking at the details of the template attack, it is important to understand the statistics concepts that are involved. A template is effectively a multivariate distribution that describes several key samples in the power traces. This section will describe what a multivariate distribution is and how it can be used in this context. Noise Distributions
Electrical signals are inherently noisy. Any time we take a voltage measurement, we don’t expect to see a perfect, constant level. For example, if we attached a multimeter to a 5 V source and took 4 measurements, we might expect to see a data set like (4.95, 5.01, 5.06, 4.98). One way of modelling this voltage source is: X = X + N where X is the noisefree level and N is the additional noise. In our example, X would be exactly 5 V. Then, N is a random variable: every time we take a measurement, we can expect to see a different value. Note that X and N are bolded to show that they are random variables. A simple model for these random variables uses a Gaussian distribution (read: a bell curve). The probability density function (PDF) of a Gaussian distribution is f(x) = $\frac{1}{\sigma \sqrt{2\pi}} e^{(x  \mu)^2 / 2\sigma^2}$ where μ is the mean and σ is the standard deviation. For instance, our voltage source might have a mean of 5 and a standard deviation of 0.5, making the PDF look like:
We can use the PDF to calculate how likely a certain measurement is. Using this distribution, f(5.1) ≈ 0.7821 f(7.0) ≈ 0.0003 so we’re very unlikely to see a reading of 7 V. We’ll use this to our advantage in this attack: if f(x) is very small for one of our subkey guesses, it’s probably a wrong guess.
Multivariate Statistics The 1variable Gaussian distribution works well for one measurement. What if we’re working with more than one random variable? Suppose we’re measuring two voltages that have some amount of noise on them. We’ll call them X and Y. As a first attempt, we could write down a model for X using a normal distribution and a separate model for Y using a different distribution. However, this might not always make sense. If we write two separate distributions, what we’re saying is that the two variables are independent: when X goes up, there’s no guarantee that Y will follow it. Multivariate distributions let us model multiple random variables that may or may not be correlated. In a multivariate distribution, instead of writing down a single variance σ, we keep track of a whole matrix of covariances. For example, to model three random variables (X, Y, Z), this matrix would be
Also, note that this distribution needs to have a mean for each random variable:
The PDF of this distribution is more complicated: The equation for k random variables is:
Creating the Template
A template is a set of probability distributions that describe what the power traces look like for many different keys. Effectively, a template says: “If you’re going to use key k, your power trace will look like the distribution f_{k}(x)”
. We can use this information to find subtle differences between power traces and to make very good key guesses for a single power trace.
Number of Traces
One of the downsides of template attacks is that they require a great number of traces to be preprocessed before the attack can begin. This is mainly for statistical reasons. In order to come up with a good distribution to model the power traces for every key, we need a large number of traces for every key. For example, if we’re going to attack a single subkey of AES128, then we need to create 256 power consumption models (one for every number from 0 to 255). In order to get enough data to make good models, we need tens of thousands of traces.
Note that we don’t have to model every single key. One good alternative is to model a sensitive part of the algorithm, like the substitution box in AES. We can get away with a much smaller number of traces here; if we make a model for every possible Hamming weight, then we would end up with 9 models, which is an order of magnitude smaller. However, then we can’t recover the key from a single attack trace  we need more information to recover the secret key.
Points of Interest
Our goal is to create a multivariate probability describing the power traces for every possible key. If we modeled the entire power trace this way (with, say, 3000 samples), then we would need a 3000dimension distribution. This is insane, so we’ll find an alternative.
Thankfully, not every point on the power trace is important to us. There are two main reasons for this:

We might be taking more than one sample per clock cycle. There’s no real reason to use all of these samples  we can get just as much information from a single sample at the right time.

Our choice of key doesn’t affect the entire power trace. It’s likely that the subkeys only influence the power consumption at a few critical times. If we can pick these important times, then we can ignore most of the samples.
These two points mean that we can usually live with a handful (35) of points of interest. If we can pick out good points and write down a model using these samples, then we can use a 3D or 5D distribution  a great improvement over the original 3000D model.
Analyzing the Data
Suppose that we’ve picked I points of interest, which are at samples s_{i}(0 ≤ i < I). Then, our goal is to find a mean and covariance matrix for every operation (every choice of subkey or intermediate Hamming weight). Let’s say that there are K of these operations (maybe 256 subkeys or 9 possible Hamming weights).
For now, we’ll look at a single operation k (0 ≤ k < K). The steps are:

Find every power trace t that falls under the category of “operation k”. (ex: find every power trace where we used a subkey of 0x01.) We’ll say that there are T_{k} of these, so t_{j, si} means the value at trace j and POI i.

Find the average power μ_{i} at every point of interest. This calculation will look like:

Find the variance v_{i} of the power at each point of interest. One way of calculating this is:

Find the covariance c_{i, i*} between the power at every pair of POIs (iandi^{*}). One way of calculating this is:

Put together the mean and covariance matrices as:
These steps must be done for every operation k. At the end of this preprocessing, we’ll have K mean and covariance matrices, modelling each of the K different operations that the target can do.
Attack Time
With a template in hand, we can finish our attack. For the attack, we need a smaller number of traces  we’ll say that we have A traces. The sample values will be labeled a_{j, si}(1 ≤ j ≤ A). First, let’s apply the template to a single trace. Our job is to decide how likely all of our key guesses are. We need to do the following:

Put our trace values at the POIs into a vector. This vector will be:

Calculate the PDF for every key guess and save these for later. This might look like:

Repeat these two steps for all of the attack traces
This process gives us an array of p_{k, j}, which says: “Looking at trace j, how likely is it that key k is the correct one?”
Combining the Results
The very last step is to combine our p_{k, j} values to decide which key is the best fit. The easiest way to do this is to combine them as:
practical template attacks
In this section we will show a differant ways to select the most important point of a power trace, that will lead us to improved computation time and make template attack more practical.
sum of difference

For every operation k and every sample i, find the average power M_{k, i}. For instance, if there are T_{k} traces where we performed operation k, then this average is

After finding all of the means, calculate all of their absolute pairwise differences. Add these up. This will give one “trace” which has peaks where the samples are usually different. The calculation looks like
An example of this sum of differences is:
Preprocessing
In practical sidechannel analysis, the raw input data is often preprocessed. Sometimes this is just due to simplicity or efficiency reasons, e. g. summarizing sampled points. There are however cases where the preprocessing step heavily affects the results. Even if no thinkable transformation can add additional information to a signal, information extraction procedures do improve. The template attack under consideration is such a case and a lucrative preprocessing transformation is described subsequently. It turns out that the transformation of the input traces from the time domain into the frequency domain is such a lucrative transformation. In our practical work, an FFT algorithm was used to accomplish this transformation (a fast algorithm to calculate the discrete Fourier transform, for background information refer to [BP85]). In order to show the impact of this preprocessing step a number of experiments were carried out. First some characteristic differences between time domain analysis and frequency domain analysis are illustrated. Afterwards, to highlight the influence of the number of selected points on the classification performance in the frequency domain, a number of experiments were carried out. After preprocessing, the resulting traces can be used to perform a template attack in exact the same way as without preprocessing. There is however a difference in the number of points to consider. Figure 6 shows the classifications results as a function of the number of selected points after preprocessing. The considered numbers of points are ranging between 1 and 40. Additionally three different minimum distances where chosen. Results show that much less points are sufficient in comparison to a template attack without the preprocessing step. At the price of performing an FFT on every input trace (those used to build up the templates as well as those to classify) we get a major advantage
Amplified Template Attack
Even if the aim of a template attack is to recover the secret key using a single trace, in many real world settings implementations allow for several iterations of the same operation with the same secret key. The application of template attacks is not restricted to stream ciphers like RC4 and can be applied to block ciphers as well. Since every symmetric cipher contains some sort of key scheduling mechanism which processes the secret key, this generalization is possible. Smartcards often use block ciphers for encryption or authentication, hence let us consider the following example: A malicious petrol station tenant, named Eve, is using a modified smartcard based payment terminal. Everytime a customer uses this terminal, Eve captures one trace of sidechannel information. This single trace could already be used by Eve to carry out a template attack. However, some customers are coming again and Eve gets hold of another trace. The template attack can easily be extended to take advantage of such situations, e.g. by adding up noiseprobabilities p(Ni) of every captured trace and applying the maximumlikelihood approach on these sums. As a consequence, the power of the attacker is amplified. Using this approach, if n is the number of iterations, the error probability of template classification is reduced by the factor √n.
Related work

Power analysis attacks on the AES  This paper describes an attack attack on the AES algorithm using the Difference of Means technique was carried out on a FieldProgrammable Gate Array (FPGA) board. The work of Coron et al. described the idea of using the Weight Power Model and experimentation was conducted on a smart card chip. Brier et al. expanded the work on CPA by proposing the use of a Hamming distance power model in place of Hamming weight while results were presented from data gathered from attacking an 8bit chip against AES. Alioto et al. proposed a novel CPA attack named Leakage Power Attack which is derived from the theory behind the Hamming Weight Power Model and results were conducted against a MC74ACT273N chip. Lastly, Mestiri et al. describes an attack on a SASEBOGII board against the AES algorithm with the goal of comparing the Hamming distance model against another derivative of the CPA called the Switching Distance model power analysis (DPA) and correlation power analysis (CPA).

Differential Power Analysis in AES: A Crypto Anatomy  The initial papers on DPA are abstract, and recent papers propose countermeasures rather than describe the attack methodology itself. None of the previous papers considered the pipeline effects in a processor, and almost all current processors contain pipelines. And most of the papers do not clearly describe the method to locate the necessary power magnitudes (corresponding to the actual instructions which are executed during the SBOX lookup) from a long power profile. This journal describes the attack in a step by step manner, so that effective countermeasures can be proposed by a larger number of researchers. Even the recent papers in DPA approach tend to assume that the reader understands the anatomy of DPA well. Yet, a number of researchers have asked us about how it is done, and this is an effort to make the steps clear. We also look at the effect the pipeline has on the attack, and methods to identify which instructions are most vulnerable to attack.

Power Analysis Based Side Channel Attack  Work related to circuit level hardware countermeasures are very few. Sprunk in his invention uses a clock that outputs a stream of random clock pulses. When the clock is unpredictable, the moment at which a certain instruction would run is also unpredictable. Therefore the obtained power traces would be misaligned. The misalignment of power traces makes it necessary to collect and analyse large number of power traces which makes the attack more time consuming. But many communication protocols such as USB and RS232 need a stable clock and therefore usage of an unpredictable clock is a disadvantage is such situations Power analysis and Testbeds.

One trace is all it takes: Machine Learningbased Sidechannel Attack on EdDSA 
The paper proposes a Convolutional Neural Network based profiling attack on the implementation of the Ed25519 Digital Signature Algorithm in the WolfSSL library on a 32 bit STM32F4 microcontroller. The attack focuses on the optimized implementation of Elliptic Curve scalar multiplication by Bernstein et al that splits the computation to work on a single nibble of the ephemeral key at a time and uses a lookup table with all possible partial multiplication results. The lookup table access is vulnerable to side channel data leakage via power analysis, similarly to the way an SBOX lookup is exploited for side channel attacks on AES. The authors evaluate the standard template attack method as well as SVM and RandomForest based attack and finally  a CNN based attack that uses a CNN architecture similar to the wellknown VGG model used for computer vision tasks, modified to work with 1D instead of 2D convolutions and fine tuned for SCA in terms of the number of layers and layer sizes. To evaluate the attacks, the authors construct a dataset mapping power traces to the respective ephemeral key nibbles processed when the power trace was recorded. Around 5K profiling (training) traces and 1K attack (test) traces are recorded and used for the evaluation. The results show that the CNN based method has a 100% accuracy of identifying the correct ephemeral key nibble given a trace from the attack set, whereas the SVM and Random Forest methods show a slighly lower accuracy and the template attack lags behind them all. As a consequence, the CNN based method is able to iteratively recover the entire 256bit ephemeral key using just a single guess of each of its 64 nibbles and the private scalar can be computed from the ephemeral key and the respective signed message using a straightforward mathematical formula, thus giving an attack the ability to forge message signatures. Another important result is that as few as 500 training traces are sufficient to reach the 100% accuracy with the CNN based attack. 
Current Events: Identifying Webpages by Tapping the Electrical Outlet 
Computers plugged into power outlets leak information by drawing variable amounts of power when performing different tasks. This work examines the extent to which this side channel leaks private information about web browsing. Characterizing the AC power side channel may help lead to practical countermeasures that protect user privacy from untrusted power infrastructure. Web browsers increasingly take advantage of hardware acceleration to implement rich user interfaces. In practice, several challenges complicate the task of identifying webpages:

Competing websites often imitate one another, resulting in similar resourceuse patterns.

Most websites optimize for load time, so many popular webpages load in approximately the same amount of time.

A complex software stack sits between a webpage and the power supply, introducing layers of indirection that may buffer or hide finegrained events.

Many websites change frequently, often presenting different content (e.g., ads) to different users.
This paper experimentally demonstrates that, despite the above challenges, different webpages do induce different powerconsumption patterns that can be used to identify them. To do so, they used a Fourier transform of the power traces to represent the entire trace in the frequency domain. Then, they built an SVM classifier to distinguish between 50 different webpages and showed it can acheive 87% precision and 74% recall.
