论文笔记：Non-Profiled Deep Learning-based Side-Channel attacks with Sensitivity Analysis（DDLA）

论文笔记：Non-Profiled Deep Learning-based Side-Channel attacks with Sensitivity Analysis（DDLA）

Benjamin Timon
eShard, Singapore

基础概念

Non-profiling attacks：
假设攻击者只能从目标设备获取跟踪。例如：
Differential Power Analysis (DPA), Correlation Power Analysis (CPA) , or Mutual Information Analysis (MIA).

Profiling attacks：
假定攻击者拥有与目标设备相同的可编程设备。例如：
Template Attacks, Stochastic attacks or Machine-Learning-based attacks.
1.在分析阶段：用收集的侧通道迹，对所有可能密钥值k∈K进行泄漏分析。
2.攻击阶段：基于泄漏分析对侧通道迹进行分类，恢复密钥值k^∗。

points of interest（POI）： 兴趣点，即侧通道迹中的泄漏点，用POI进行分类。

Deep-Learning（DL）： MLP,CNN

de-synchronized： 去同步，未对齐

Contribution

• 提出Differential Deep Learning Analysis (DDLA) .
• 着重实现Sensitivity Analysis in a Non-Profiled context.

DDLA

攻击算法
AES

DDLA具体算法

实验硬件&框架

• PC:64 GB of RAM, a GeForce GTX 1080Ti GPU & two Intel Xeon E5-2620 v4 @2.1GHz CPUs.
• ChipWhisperer-Lite(CW) ；Atmel XMEGA128 chip;
  ASCAD(collected from an 8-bit ATMega8515 board).
• MLP_exp & CNN_exp.

实验参数
1 Trainings parameters and details

1.1 Loss function
	Mean Squared Error (MSE) loss function for all experiments.
1.2 Accuracy
	The accuracy was computed as the proportion of samples correctly classified.
1.3 Batch size
	A batch size of 1000 was used for all experiments.
1.4 Learning rate
	For all experiements we used a learning rate of 0.001.
1.5 Optimizer
	We used the Adam optimizer with default configuration (β~1~ = 0.9, β~2~ = 0.999, ε = 1e-08,no learning rate decay).
1.6 Input normalization
	We normalize the input traces by removing the mean of the traces and scaling the traces between -1 and 1.
1.7 Labeling
	• For all simulations (first and high-order), we used the MSB labeling.
	• For attacks on the unprotected CW and on ASCAD, we used the LSB labeling.
	• For the attack on the CW with 2 masks, we used the MSB labeling.
1.8 Deep Learning Framework
	We used PyTorch 0.4.1.

PyTorch

2 Networks architectures
Deeping Learning Book
Deep learning website

2.1 MLPsim
	• Dense hidden layer of 70 neurons with relu activation
	• Dense hidden layer of 50 neurons with relu activation
	• Dense output layer of 2 neurons with softmax activation
2.2 CNNsim
	• Convolution layer with 8 filters of size 8 (stride of 1, no padding) with relu activation.
	• Max pooling layer with pooling size of 2.
	• Convolution layer with 4 filters of size 4 (stride of 1, no padding) with relu activation.
	• Max pooling layer with pooling size of 2.
	• Dense output layer of 2 neurons with softmax activation

2.3 MLPexp
	• Dense hidden layer of 20 neurons with relu activation
	
	• Dense hidden layer of 10 neurons with relu activation
	• Dense output layer of 2 neurons with softmax activation
2.4 CNNexp
	• Convolution layer with 4 filters of size 32 (stride of 1, no padding) with relu activation.
	• Average pooling layer with pooling size of 2.
	• Batch normalization layer
	
	• Convolution layer with 4 filters of size 16 (stride of 1, no padding) with relu activation.
	• Average pooling layer with pooling size of 4.
	• Batch normalization layer
	• Dense output layer of 2 neurons with softmax activation

Experiment & result

CNN-DDLA against de-synchronized traces

Non-Profiled；N = 3, 000 未对齐侧信道迹；CNN_exp .
（文章用了MLP_exp和CPA，未能成功恢复）

High-Order DDLA simulations （掩码技术泄露处未知）： MLP_sim

N = 5, 000 traces，1 mask ；
• n = 50 samples per trace.
• leakage： t = 25 ；Sbox leakage：Sbox(di ⊕ k^∗) ⊕ m1 + N (0, 1) ，di & m1 are random | k^∗ a fixed key byte.
leakage：t = 5 and defined as m1 + N (0, 1).
• All other points on the traces are random values in [0; 255]

N = 10, 000 traces， 2 masks.；
leakage：t = 45；Sbox leakage：Sbox(di ⊕ k∗) ⊕ m1 ⊕ m2 + N (0, 1).

Second order DDLA on ASCAD

16-byte fixed key；plaintexts & masks are random；ASCAD profiling set；MLP_exp on the first 20, 000 traces；n_e = 50 epochs per guess

Third order DDLA on ChipWhisperer

N = 50, 000 traces；n = 150 samples with mask m1,m2；masked Sbox：Sbox(d ⊕ k∗) ⊕ m1 ⊕ m2.
MLP_exp network；n_e = 100 epochs per guess；
reveal key after around 20 epochs per guess；without any leakages combination pre-processing nor any assumptions about the masking method.

这些是用DL训练每个bit的密钥，最后检查矩阵；
但是以某部分epochs为单位检查矩阵，直到某一步骤可以恢复密钥，可以减少复杂度。