TAME: A Deep Learning-based Blind Side-Channel Analysis Framework

Abstract

Side-Channel Analysis (SCA) methods critically depend on access to known plaintext/ciphertext, an assumption violated in many practical scenarios such as OCB encryption mode and the EMV payment protocol. Blind SCA addresses this challenge by operating without plaintext/ciphertext, relying instead on a three-phase pipeline whose effectiveness hinges on the quality of trace labeling methods. Existing deep learning-based Blind SCA (DL-BSCA) method achieves promising performance but still suffers from two fundamental limitations: (1) multivariate clustering induces an excessively large joint Hamming weight space with exponential growth which degrades DL-BSCA’s performance due to the curse of dimensionality; and (2) all labeled traces are treated with uniform weight during training, ignoring the class information implied in the geometric space, such as the difference in labeling accuracy between samples close to cluster centers and those near cluster boundaries.To address both limitations, we propose a new DL-BSCA framework comprising two mutually reinforcing components. First, we introduce Multi-point Cluster Deep R epresentation (MCDR) labeling, which performs clustering in a compact learned feature space rather than the raw Hamming weight space, mitigating noise sensitivity and dimensionality explosion. Second, we propose Trust Anchor Meta-weight Estimation (TAME), a bilevel optimization framework that constructs a high-confidence validation set from cluster-interior traces, termed Trust Anchor (TA) set, and automatically learns per-sample weights guided by validation loss, effectively suppressing the influence of mislabeled traces. Better cluster separation from MCDR yields higher-quality TA set, which in turn strengthens TAME’s meta-weight supervision. Experiments target three standard cryptographic algorithms, namely AES, ASCON, and Kyber, which demonstrate that our approach recovers the secret key using at most 8000 traces, while prior methods fail with guessing entropy consistently at least 3. Under desynchronization countermeasures, our method succeeds with as few as 150 traces where all existing approaches fail to reach a guessing entropy equal to zero, highlighting its robustness in challenging real-world conditions.