Machine Learning Power Side-Channel Attack Exposes Weakness in SNOW‑V Encryption

Global: Machine Learning Power Side-Channel Attack Exposes Weakness in SNOW‑V Encryption

Researchers Deepak, Rahul Balout, Anupam Golder, Suparna Kundu, Angshuman Karmakar and Debayan Das submitted a paper on 25 December 2025 describing a power‑analysis side‑channel attack against the SNOW‑V stream cipher, a candidate for 5G mobile‑communication security. The experiment was performed on an STM32 microcontroller, with power traces captured by a ChipWhisperer board, and the findings were reported in the arXiv preprint arXiv:2512.21737.

Methodology

The authors applied Test Vector Leakage Assessment (TVLA) to confirm the presence of exploitable leakage in the device’s power consumption. Subsequent profiling attacks leveraged both Linear Discriminant Analysis (LDA) and Fully Connected Neural Networks (FCN) to model the relationship between power traces and secret key material.

Machine‑Learning Models

While LDA provided a baseline for key recovery, the FCN approach achieved a markedly higher efficiency. According to the study, the neural‑network model required more than five times fewer traces to reach disclosure compared with the traditional Correlational Power Analysis (CPA) method assisted by LDA.

Results and Comparison

The minimum traces to disclosure (MTD) metric demonstrated that the FCN‑based attack reduced the trace count needed for successful key extraction, establishing a new performance benchmark for side‑channel attacks on SNOW‑V.

Implications for 5G Security

The results suggest that SNOW‑V, despite being a promising 5G encryption standard, is vulnerable to modern machine‑learning‑driven side‑channel techniques. This raises concerns for any deployment that relies on the algorithm without additional hardware or software countermeasures.

Recommendations and Future Work

The authors call for the development of robust countermeasures, such as masking or hiding techniques, and recommend further investigation into alternative machine‑learning models and hardware platforms to fully assess the attack surface.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via arXiv.