The authors model an adversary with write access to model weights before deployment, such as a malicious insider or a compromised supply chain.
Backdoors can hide executable neural networks inside model weights
Draguns et al. embed encrypted capabilities behind secret triggers and show they resist fine-tuning, noising, trigger search, and adversarial training.
Top University
Andis Draguns · Sumeet Ramesh Motwani · Raymond Douglas · Christian Schroeder de Witt
Contramont Research · University of Oxford · ACS Research
Research Digest··3 min read
Draguns et al.
Why this paper
From University of Oxford and 2 others
In one line
Backdoor attacks can embed encrypted neural networks in model weights, hiding arbitrary capabilities from elicitation even with full white-box access.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§