Evaluating Representation Learning for Profiling Side-Channel Forensics in Resource-Constrained IoT Devices

Deo N, Fischer G (2026)

Publication Language: English

Publication Type: Conference contribution, Abstract of a poster

Publication year: 2026

City/Town: Linköping

Event location: Linköping SE

URI: https://dfrws.org/conferences/dfrws-eu-2026/

Abstract

Side-channel attacks (SCA) exploit physical emanations —power consumption, electromagnetic radiation, timing —rather than mathematical weaknesses. On resource-constrained IoT and edge AI devices, SCA is particularly potent because masking is expensive and physical isolation is absent. Side-channel attacks traditionally break cryptography by correlating physical signals (power, EM emissions) with intermediate values. Recent work shows EM side-channels can detect IoT firmware modifications and cryptographic operations with high accuracy . Thus, SCA techniques could be integrated into forensic toolkits for IoT, but they must be adapted to device constraints. The threat is quantified: Benadjila et al. (2020) showed MLP trained on 50K power traces of a masked ATmega8515 can reduce key rank to near-zero. Zaid et al. (2019) demonstrated CNNs achieving GE≈1 on the ASCAD dataset in under 1,000 traces —meaning the full 128-bit key is essentially recovered. Profiling attacks assume the adversary has a cloned device: they build a leakage model offline (profiling phase), then apply it to the target with very few traces (attack phase). This matches real forensic scenarios —seized device models are publicly available or purchasable.

Authors with CRIS profile

Navneeta Deo Professur für Technische Elektronik Georg Fischer Professur für Technische Elektronik

How to cite

APA:

Deo, N., & Fischer, G. (2026, March). Evaluating Representation Learning for Profiling Side-Channel Forensics in Resource-Constrained IoT Devices. Poster presentation at The Digital Forensics Research Conference Europe (DFRWS EU 2026), Linköping, SE.

MLA:

Deo, Navneeta, and Georg Fischer. "Evaluating Representation Learning for Profiling Side-Channel Forensics in Resource-Constrained IoT Devices." Presented at The Digital Forensics Research Conference Europe (DFRWS EU 2026), Linköping Linköping, 2026.

BibTeX: Download