Year: 2026
Type: article
Abstract: This paper introduces Heuriskryptes, a detection framework designed to identify AI-driven steganographic attacks within deep neural networks. The framework addresses the emerging threat of covert channels established within a model’s parameter pool. The evaluated attack vector, termed Stego-GCP, uses an orthogonal embedding to modify model weights with high stealth to leak data. To detect these channels, a specialized agent utilizes a Vision Transformer (ViT) architecture that treats the raw parameter space as an image, providing a self-attention-based mechanism that demonstrates consistent cross-architecture generalizability for efficient model auditing. The framework has been validated using a PooledCNN architecture on the Apple Disease and CIFAR-10 datasets. Results indicate that the carrier model retains a classification accuracy of 99.97%. Furthermore, the Heuriskryptes detector demonstrated high sensitivity, achieving 93.51% overall accuracy and maintaining a 90% detection rate specifically on the most challenging ambiguous, low-payload density scenarios ( $$\alpha < 0.002$$ ). This demonstrates highly competitive performance compared to evaluated frequency-domain anomaly detection baselines. Heuriskryptes serves as a reliable detection framework for verifying model integrity in collaborative and commercial AI environments. By focusing on the structural mathematical properties of the weight spectrum rather than activation-trigger patterns, this research provides a reliable defense against the unauthorized use of neural architectures as covert communication channels.
Source: Scientific Reports
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7231427697
