Protection of devices against side channel attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing countermeasures against side-channel attacks, such as deep learning models, are vulnerable to attacks that exploit identical data distributions between training and inference devices, requiring impractical computation to overcome.
Innovation Solution
A method that modifies the execution of sensitive algorithms by generating signal modification data based on a unique device secret, altering the data distribution to render training models ineffective on inference devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If classical countermeasures such as random mask values and shuffling are used, then side-channel attacks are partially mitigated, but machine-learning models can still bypass these countermeasures by learning to recognize and detect them during training and inference
Solution Approach 1:
The patent applies dynamics by making the countermeasures adaptive and variable rather than static. Each device generates unique countermeasures based on its specific characteristics (probe position, chip thinning, oscilloscope setup), and these countermeasures change dynamically during execution. This prevents machine-learning models from learning fixed patterns, as the same algorithm execution produces different trace characteristics on different devices.
Solution Approach 2:
The patent changes the parameters of the executed algorithm by introducing device-specific modifications to the sensitive algorithm execution. These parameter changes include modifying the algorithm's behavior based on device characteristics, which alters the leakage signal characteristics. This makes the data distribution on different devices non-identical, rendering cross-device attacks ineffective.
2Measurement precision
If deep learning models are trained on traces from a single device with specific setup, then the model can accurately recover secrets, but ensuring the same data distribution on a different inference device requires exact replica of the setup which is challenging and complex
Solution Approach 1:
The patent introduces asymmetry by ensuring that each device has unique countermeasures tailored to its specific characteristics. Instead of attempting to create identical symmetric conditions across devices, the system embraces the differences by generating device-specific countermeasures. This asymmetric approach makes it impossible for an attacker to train a model on one device and successfully apply it to another, as the leakage patterns will differ due to the asymmetric device characteristics and corresponding countermeasures.
3Loss of information
If the entire trace is used for training deep learning models, then more information is available for secret recovery, but the computation requirements become impractical without pre-processing steps
Solution Approach 1:
The patent extracts only the necessary information by making the countermeasures cause the sensitive algorithm to execute at different times or in different ways on different devices. This temporal and execution variation effectively removes the need for extensive pre-processing and manual feature extraction, as the device-specific countermeasures automatically create distinguishable leakage patterns that can be directly used for attack detection without requiring complex pre-processing to identify relevant features.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
A computer-implemented method of modifying the execution of a sensitive algorithm by a device having a secret S associated with it, wherein execution of the sensitive algorithm comprises using a critical value ki. The computer-implemented method comprises: generating signal modification data based on the secret S, the signal modification data defining one or more countermeasures to be executed during execution of the sensitive algorithm; and executing the sensitive algorithm and one or more countermeasures according to the signal modification data.