# From Capture–Recapture to No Recapture: Efficient SCAD Even After Software Updates

## TL;DR

This paper introduces a new method to efficiently detect unauthorized software changes in IoT devices using synthetic electromagnetic signals, reducing the need for manual updates.

## Contribution

The novel contribution is a generative modeling framework using CWGAN-GP with ESD conditioning to synthesize EM signals for SCAD after software updates.

## Key findings

- The proposed framework generates synthetic EM signals with 85–92% similarity to real signals at instruction-level granularity.
- ESD conditioning improves signal fidelity by reducing similarity distance by ∼13%.
- A 1DCNNGAN model outperforms ResGAN in training speed and memory efficiency while maintaining detection accuracy.

## Abstract

Side-Channel-based Anomaly Detection (SCAD) offers a powerful and non-intrusive means of detecting unauthorized behavior in IoT and cyber–physical systems. It leverages signals that emerge from physical activity—such as electromagnetic (EM) emissions or power consumption traces—as passive indicators of software execution integrity. This capability is particularly critical in IoT/IIoT environments, where large fleets of deployed devices are at heightened risk of firmware tampering, malicious code injection, and stealthy post-deployment compromise. However, its deployment remains constrained by the costly and time-consuming need to re-fingerprint whenever a program is updated or modified, as fingerprinting involves a precision-intensive manual capturing process for each execution path. To address this challenge, we propose a generative modeling framework that synthesizes realistic EM signals for newly introduced or updated execution paths. Our approach utilizes a Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) framework trained on real EM traces that are conditioned on Execution State Descriptors (ESDs) that encode instruction sequences, operands, and register values. Comprehensive evaluations at instruction-level granularity demonstrate that our approach generates synthetic signals that faithfully reproduce the distinctive features of real EM emissions—achieving 85–92% similarity to real emanations. The inclusion of ESD conditioning further improves fidelity, reducing the similarity distance by ∼13%. To gauge SCAD utility, we train a basic semi-supervised detector on the synthetic signals and find ROC-AUC results within ±1% of detectors trained on real EM data across varying noise conditions. Furthermore, the proposed 1DCNNGAN model (a CWGAN-GP variant) achieves faster training and reduced memory requirements compared with the previously leading ResGAN.