Study tests black-box watermarks for hyperdimensional classifiers
Original title: Intellectual Property Protection of Hyperdimensional Computing Classifiers: Black-Box Ownership Verification via Model Watermarking
AdaptHD and OnlineHD kept verification at 1% trigger exposure with at most 0.58 points of mean accuracy loss.
Researchers study backdoor-style watermarking for hyperdimensional computing classifiers to verify model ownership. Across four benchmarks, error-driven AdaptHD and OnlineHD supported verification at a 1% trigger exposure with at most 0.58 percentage points of mean accuracy loss, while single-pass Vanilla HDC needed more exposure and could lose more accuracy. Extraction surrogates copied task behavior without verifiable watermarks, and overwriting damaged owner marks more than in neural-network references. FPGA tests of watermarked AdaptHD survived some undervolting and overclocking errors.
Unconfirmed: Retention, overwriting, and hardware-stress outcomes are limited to the authors' benchmarks and FPGA tests.
- Classifiers studied are AdaptHD, OnlineHD, and Vanilla HDC.
- Nominal trigger exposure for the error-driven models is 1%.
- Mean accuracy loss was at most 0.58 percentage points in that setting.
- Extraction removed verifiable watermarks from surrogate models.
- Severe FPGA stress caused verification failure, accuracy collapse, or timeouts.
- Products
- AdaptHDOnlineHDVanilla HDC
From the original article
1 min read at arxiv.orgHyperdimensional computing (HDC) offers efficient learning and inference for resource-constrained systems, yet ownership verification of its models remains underexplored. We present the first systematic investigation of model watermarking in HDC by adapting backdoor-based embedding to three classifiers with different prototype-learning mechanisms.
By Hamed Poursiami · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.