RealFeedApp beta
- PyTorch experiments fail to extract encrypted neural networks
By implementing encrypted neural networks within PyTorch weights, the authors prove that arbitrary computations remain hidden from fine-tuning and gradient-based search under cryptographic assumptions.
- Field
- how neural networks work
- What they did
- Researchers demonstrated that AI models can hide arbitrary computational capabilities by embedding encrypted neural networks within the host model's weights, which only execute upon receiving a secret trigger, making them provably difficult to elicit even with full access to the model's weights.
- Why it matters
- This enables the creation of systems that do not exhibit dangerous behavior under normal conditions but can perform specific tasks when a specific trigger is present, expanding the understanding of AI vulnerabilities and reliability.
#ai safety#backdoor attacks#neural networks#cryptographic assumptions#adversarial conditions
