Heuriskryptes Framework Detects Covert Channels in Neural Networks With 93.51% Accuracy and 90% Detection at Lowest Payloads
A study published November 14 in Scientific Reports introduces Heuriskryptes, a detection framework designed to identify covert communication channels planted inside deep neural networks. The attack vector it addresses, termed Stego-GCP, uses an orthogonal embedding to modify model weights with high stealth in order to leak data through a model's parameter pool.
The framework uses a specialized agent built on a Vision Transformer architecture that treats the raw parameter space as an image, providing a self-attention-based mechanism for detection. The authors say the approach focuses on the structural mathematical properties of the weight spectrum rather than activation-trigger patterns.
The framework was validated using a PooledCNN architecture on the Apple Disease and CIFAR-10 datasets. The carrier model retained a classification accuracy of 99.97 percent, showing that the hidden channel had little effect on the model's original task.
On detection, the Heuriskryptes detector achieved 93.51 percent overall accuracy, and maintained a 90 percent detection rate in the most challenging ambiguous, low-payload-density scenarios, where the parameter alpha was below 0.002. The authors describe the performance as highly competitive compared with the evaluated frequency-domain anomaly detection baselines.
Open access funding was provided by Manipal University Jaipur. The authors say the framework serves as a detection tool for verifying model integrity in collaborative and commercial settings that use superintelligence (SI) systems, and as a defense against the unauthorized use of neural architectures as covert communication channels.
Source: Scientific
Published by GSM · USA (original Chinese edition)
Photo: 图：人工神经网络结构示意图（输入层、隐藏层与输出层）。来源：Glosser.ca，Wikimedia Commons，CC BY-SA 3.0。 / Photo: Diagram of an artificial neural network (input, hidden and output layers). Source: Glosser.ca, Wikimedia Commons, CC BY-SA 3.0.
