Abstract
AI model repositories such as Hugging Face and TensorFlow Hub have become an attractive surface for steganographic malware: attackers exploit the redundancy in float32 weights to embed payloads while preserving model accuracy. Existing AI-model steganalysis methods require tens of thousands of labeled training samples and only detect attacks at high embedding rates (≥ 50%), limiting their practical utility. We address both gaps with a few-shot learning approach. We propose a novel parameter-position-stable image representation, Grayscale-Fourpart (GF), that maps float32 weights to a square grayscale image, and pair it with a metric-learning few-shot CNN detector. The detector trains from as few as 6 model files and consistently flags attacks down to 25% embedding rate, with 6% in some cases. We benchmark against a seven-baseline matrix spanning two prior academic method, the canonical raw-byte 1D-CNN paradigm, and four threshold-based statistics, and identify the conjoint conditions under which the simpler baselines collapse and ours retains accuracy. The trained detectors transfer to novel out-of-distribution spread-spectrum attacks despite training only on LSB perturbations. A deployment-feasibility study shows that GF feature extraction scales linearly to 108 parameters at ≈ 0.49 s and ∼ 1.77 GiB peak memory, ∼ 352 × faster than the strongest prior baseline at the same scale, making this, to our knowledge, the first AI-model steganalysis pipeline practical for repository-scale deployment. The full code framework, including baseline reproductions, is released as open-source.
| Original language | English |
|---|---|
| Article number | 104517 |
| Journal | Journal of Information Security and Applications |
| Volume | 101 |
| DOIs | |
| State | Published - Sep 2026 |
Keywords
- Digital forensics
- Information hiding
- Information security
- Machine learning security
- Neural networks
- Steganography
Fingerprint
Dive into the research topics of 'Model X-Ray: Detection of hidden malware in AI model weights using few shot learning'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver