Breaking the Blindfold: Deep Learning-based Blind Side-channel Analysis

Azade Rezaeezade, Trevor Yap, Dirmanto Jap, Shivam Bhasin, Stjepan Picek

TL;DR

This paper introduces DL-BSCA, a novel framework that uses deep learning to perform blind side-channel analysis (SCA), a challenging attack scenario where attackers have no knowledge of plaintexts or ciphertexts. By developing a new labeling technique and leveraging deep neural networks, they achieve practical key recovery on various cryptographic algorithms and hardware platforms, even against countermeasures that previously rendered such attacks impossible.

What Problem Does This Paper Address?

Physical side-channel attacks (SCA) are a significant threat to the security of cryptographic implementations. Traditional SCA methods typically rely on having access to known plaintexts or ciphertexts to extract secret keys from physical leakages like power consumption. However, in many real-world scenarios, such as certain encryption modes (e.g., OCB) or complex hardware implementations, this "known plaintext/ciphertext" assumption is invalid. This leads to the "blind SCA" problem, where attackers must deduce the secret key solely from side-channel traces without any knowledge of the input or output data. While prior blind SCA approaches exist, they have shown limited success, often requiring high signal-to-noise ratio (SNR) devices and failing in practical settings or against countermeasures. This paper aims to bridge this gap by developing a more robust and practical blind SCA framework.

Methods

The core of this work is the Deep Learning-based Blind Side-channel Analysis (DL-BSCA) framework. It extends the traditional three-phase blind SCA process:

  1. Phase 1: Pre-computing theoretical joint distribution: This phase remains standard, where theoretical distributions of Hamming weights for intermediate values are computed for all possible keys.
  2. Phase 2: Labeling traces to obtain empirical distribution: This is where the main innovation lies. Instead of relying on simple slicing or variance analysis on a single "point of interest" (PoI) per variable, the authors propose:
  3. Multi-point Cluster-based (MC) labeling: This novel technique selects multiple PoIs (e.g., 50 for each variable) to better represent the leakage. These PoIs are then fed into a Gaussian Mixture Model (GMM) clustering algorithm to group traces with similar leakage patterns. The clusters are then labeled using a weighted majority voting scheme to assign a single label (Hamming weight tuple) to each cluster. This approach accounts for dependencies between variables and is more robust to noise.
  4. Deep Learning for Labeling: DL-BSCA treats blind SCA as a deep learning problem with noisy labels. The MC labeling is used to label a subset of traces (training set) to train a deep neural network (DNN). The trained DNN then predicts labels for the remaining traces (attack set), providing the empirical distribution. This leverages the DNN's ability to generalize even with imperfect labels.
  5. Phase 3: Comparing empirical and theoretical distributions: This phase uses the maximum likelihood criterion to compare the distributions and extract the secret key, which is a standard approach.

The paper explores both Convolutional Neural Networks (CNNs) and Multi-Layer Perceptrons (MLPs) as the DNN architectures.

Datasets

The authors validate their DL-BSCA framework on four datasets across various platforms and cryptographic algorithms:

  • ChipWhisperer (CW) Dataset: An unprotected AES software implementation on an 8-bit AVR XMEGA microcontroller. This is a common benchmark for SCA.
  • Kyber Dataset: Based on the CRYSTAL-Kyber KEM implementation from the pqm4 library, running on a 32-bit ARM Cortex-M4 microcontroller. This is a post-quantum cryptography algorithm.
  • ASCON Dataset: An implementation of the ASCON lightweight cryptography algorithm, also on a 32-bit ARM Cortex-M4 microcontroller.
  • Desynchronized CW Dataset: The standard CW dataset with random desynchronization applied to the traces, simulating a common countermeasure.

Key Insights and Ideas

  • Blind SCA is a Deep Learning problem with noisy labels: Framing blind SCA as a noisy-label learning problem allows leveraging the robustness of DNNs, which can generalize better than traditional methods when labels are imperfect.
  • Multi-point labeling is crucial for robustness: Moving beyond single PoIs to multiple, clustered PoIs significantly improves the ability to label traces accurately, even in the presence of noise and complex leakages.
  • MC labeling with GMM is effective: The proposed MC labeling, combined with GMM clustering, provides a more comprehensive and accurate way to assign labels in blind SCA compared to prior methods.
  • DNNs significantly boost blind SCA performance: DL-BSCA consistently outperforms classical blind SCA methods across all tested datasets and algorithms, often achieving successful key recovery where prior methods failed.
  • Practical blind SCA is achievable on diverse platforms: The framework demonstrates success on both low-cost 8-bit microcontrollers and more complex 32-bit ARM platforms, highlighting its broad applicability.
  • First successful blind SCA against desynchronization: The paper reports the first practical blind SCA against a desynchronization countermeasure, a significant advancement for blind SCA capabilities.

Key Results

  • Improved Key Recovery: DL-BSCA consistently achieved significantly lower Guessing Entropy (GE) compared to prior blind SCA methods (Linge et al. [53], Clavier & Reynaud [13]) across all tested datasets.
  • For the CW dataset, DL-BSCA with MC labeling and dropout achieved GE < 10, while prior methods failed.
  • For Kyber, DL-BSCA with MC labeling achieved GE < 10, whereas classical methods failed.
  • For ASCON, DL-BSCA with MC labeling achieved GE < 10, outperforming prior methods.
  • Desynchronized CW Dataset: DL-BSCA with MC labeling achieved GE < 4, marking the first successful blind SCA against this countermeasure.
  • Practicality Demonstrated: All experiments were validated with real-world SCA measurements, not just simulations.
  • Effectiveness of MC Labeling: The MC labeling technique, especially with 50 PoIs per variable, showed superior performance compared to Slicing or VA labeling when used with DL-BSCA.
  • Accuracy vs. GE: The paper highlights that per-trace accuracy of labeling is not a good indicator of attack success; GE is a more relevant metric, and MC labeling with DNNs achieved good GE even with low labeling accuracy.

Limitations and Future Work

  • PoI Selection: A key limitation acknowledged is that the selection of informative PoIs still requires some knowledge or correlation analysis, which is a non-blind step. Developing a fully blind PoI selection method remains an open challenge.
  • Masking Countermeasures: The current DL-BSCA framework is not directly applicable to masked implementations. Randomized shares obscure leakage, making informative labeling difficult. Attacking masked implementations in a practical blind setting is an open problem.
  • Dataset and Model Dependency: The optimal number of PoIs (e.g., 50) can be dataset and model-dependent.
  • Future Work: The authors suggest addressing the PoI selection challenge and exploring attacks on masked implementations. They also note that the MC labeling technique can be generalized to more than two variables.

Summary updated

BibTeX
@article{rezaeezade2025breaking,
  author = {Azade Rezaeezade and Trevor Yap and Dirmanto Jap and Shivam Bhasin and Stjepan Picek},
  title = {{Breaking the Blindfold: Deep Learning-based Blind Side-channel Analysis}},
  journal = {IACR Cryptol. ePrint Arch.},
  year = {2025},
}

Loading related content...

Figure