Generalized Power Attacks against Crypto Hardware using Long-Range Deep Learning (2306.07249v2)

Abstract: To make cryptographic processors more resilient against side-channel attacks, engineers have developed various countermeasures. However, the effectiveness of these countermeasures is often uncertain, as it depends on the complex interplay between software and hardware. Assessing a countermeasure's effectiveness using profiling techniques or machine learning so far requires significant expertise and effort to be adapted to new targets which makes those assessments expensive. We argue that including cost-effective automated attacks will help chip design teams to quickly evaluate their countermeasures during the development phase, paving the way to more secure chips. In this paper, we lay the foundations toward such automated system by proposing GPAM, the first deep-learning system for power side-channel analysis that generalizes across multiple cryptographic algorithms, implementations, and side-channel countermeasures without the need for manual tuning or trace preprocessing. We demonstrate GPAM's capability by successfully attacking four hardened hardware-accelerated elliptic-curve digital-signature implementations. We showcase GPAM's ability to generalize across multiple algorithms by attacking a protected AES implementation and achieving comparable performance to state-of-the-art attacks, but without manual trace curation and within a limited budget. We release our data and models as an open-source contribution to allow the community to independently replicate our results and build on them.

• The paper introduces GPAM, a novel deep-learning architecture that automates power side-channel analysis across multiple cryptographic algorithms.
• GPAM leverages temporal patchification and Transformer encoder blocks to efficiently process raw power traces without manual curation.
• Experimental evaluations demonstrate competitive success against hardened AES and ECC implementations, prompting new directions in cryptographic security research.

Analyzing Cryptographic Hardware via Generalized Power Attacks with Long-Range Deep Learning

Elie Bursztein et al. introduce GPAM (Generalized Power Analysis Model), a sophisticated deep-learning architecture tailored for power side-channel analysis on cryptographic hardware. GPAM distinguishes itself by offering a unifying approach capable of analyzing multiple cryptographic algorithms and implementations, specifically targeting AES and ECC. The authors' research addresses the limited adaptability and high expertise requirements of current machine learning-based side-channel attacks, known as SCAAMLs.

Overview and Methodology

The paper emphasizes the vulnerability of cryptographic co-processors to side-channel attacks which exploit physical signal emissions to uncover secrets such as AES keys. Traditional countermeasures have focused on specific algorithmic protections, but they often fail to provide robust defenses in the face of varying attack vectors. GPAM aims to alleviate these adaptability issues by enabling profiling-based attacks that generalize automatically across implementations and algorithms without manual intervention.

The architecture of GPAM is crafted to bypass specific limitations identified in prior efforts. Notably, GPAM utilizes temporal patchification and Transformer encoder blocks to process extensive raw power traces effectively, avoiding the expensive preprocessing and expert intervention traditionally required. This approach efficiently tackles long-range data dependencies inherent in high-fidelity cryptographic traces. The multi-task learning framework facilitates the model's adaptability, allowing it to accommodate varied cryptographic algorithms and countermeasures in its evaluations.

Experimental Evaluation

Bursztein et al. validate GPAM's efficacy through extensive experimentation, demonstrating successful attacks on numerous hardened hardware-accelerated cryptographic implementations, such as AES-protected and ECDSA implementations. The model consistently showcases competitive performance with state-of-the-art attacks but operates without the requisite manual trace curation, demonstrating the inherent advantages of generalized models.

For ECC, GPAM achieved notable success across four hardware implementations with varying levels of security, accurately predicting crucial elements of protected computations. The tests revealed that predictions using intermediate variables considerably enhanced attack success under white-box conditions, underscoring the importance of multi-task learning. Even in challenging scenarios like CM3, GPAM identified leakages that were leveraged using lattice-based analysis to recover cryptographic secrets effectively.

Implications and Future Directions

The theoretical contributions of this investigation lie in establishing a methodological foundation for automated side-channel analyses. GPAM exemplifies the potential for machine learning architectures to offer scalable solutions against multiple cryptographic challenges, preserving analytical efficiency while providing robust attack models. The open-sourcing of GPAM's models and datasets further stimulates research reproducibility and advancements in cryptographic security.

While GPAM demonstrates high adaptability and effectiveness, it raises several compelling research questions for future exploration. Key among these is the development of countermeasures that sufficiently confound such generalized attacks. Moreover, techniques that better stabilize training against initialization influences and improve upon data efficiency remain as promising future directions. Additionally, expanding GPAM's capabilities to include other cryptographic algorithms and hardware designs could augment its applicability.

Overall, GPAM marks a pivotal step in evolving cryptographic attack strategies emphasizing automatic, scalable, and cross-implementation adaptability, encouraging a shift towards versatile solutions in the security domain.