Researchers put encrypted networks in the weights of AI models
If the cryptographic assumptions hold, the circuits stay hard to find, also with full access to the weights.
Claimed, not confirmed
The researchers made backdoor attacks that put encrypted neural networks in the weights of a host model. The secret network decrypts and runs only when the model gets a secret trigger. The group examined fine-tuning, weight noising, gradient-based trigger search and latent adversarial training. All of them did not find the secret function. The group shows that audits which use elicitation have a limit.
Sources
Posted
