Tessera offers secure, near-line-rate weight streaming for edge AI accelerators

Researchers have developed Tessera, a new architecture designed to securely stream model weights to edge accelerators in Unified Memory Architecture (UMA) systems. This approach addresses the challenge of protecting proprietary deep neural networks on commodity devices by enabling inline, cache-line granularity decryption of weights. Tessera intercepts memory bursts and decrypts them in parallel with DRAM fetches, streaming plaintext directly into isolated NPU SRAM with minimal bandwidth overhead. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enhances security for deploying proprietary models on edge devices by enabling efficient, hardware-backed DRM.

RANK_REASON Academic paper detailing a new architecture for secure weight streaming on edge accelerators.

Read on arXiv cs.LG →

• Tessera
• UMA
• AES-256-CTR
• SoC
• SRAM
• DRAM
• NPU
• CPU
• DNN

• paper
• infra