TT-SEAL: TTD-Aware Selective Encryption for Adversarially-Robust and Low-Latency Edge AI

Researchers developed TT-SEAL, a selective encryption framework for compressed AI models using Tensor-Train Decomposition that maintains security while encrypting only 4.89-15.92% of parameters. The system achieves the same robustness as full encryption while reducing AES decryption overhead in end-to-end latency from 58% to as low as 2.76%.

• →TT-SEAL selectively encrypts only the most critical parameters in compressed AI models, requiring encryption of just 4.89-15.92% of total parameters.
• →The framework maintains equivalent security to full black-box encryption while dramatically reducing computational overhead.
• →AES decryption latency drops from 58% to 2.76% of total processing time on ResNet-18 models.
• →The system was successfully prototyped on FPGA hardware and tested across multiple popular neural network architectures.
• →This approach enables secure AI deployment on resource-constrained edge devices without sacrificing performance or security.