Researchers have investigated model watermarking techniques for hyperdimensional computing (HDC) classifiers, adapting backdoor-based embedding to three distinct HDC models. The study found that error-driven classifiers like AdaptHD and OnlineHD can support verification with minimal accuracy loss, while Vanilla HDC requires more exposure and can significantly impact performance. The watermarks demonstrated resilience to continued training but were vulnerable to extraction and overwriting by attacker watermarks, though severe overwriting also led to accuracy degradation. When tested on a field-programmable gate array (FPGA), the watermarked AdaptHD model showed verification survival even under substantial computational errors. AI
IMPACT This research explores methods for protecting intellectual property in efficient AI models, potentially impacting the development and deployment of AI in resource-constrained environments.
RANK_REASON The cluster contains an academic paper detailing new research findings on model watermarking for hyperdimensional computing classifiers.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
