A new technique allows attackers to exfiltrate neural network weights from machine learning models, potentially exposing proprietary AI systems. Security researchers demonstrated the vulnerability across multiple model architectures.
Security researchers have published findings on a method to extract model weights from neural networks, raising concerns about intellectual property protection in AI systems.
The technique exploits how models process queries to reconstruct their internal parameters without direct access to the underlying weights. Researchers tested the approach on various architectures and confirmed successful extraction under multiple conditions.
The vulnerability highlights risks for organizations deploying machine learning models in cloud environments or through API endpoints. Extracted weights could enable competitors to replicate proprietary models or conduct further attacks.
The discovery has sparked discussion in the security community about defenses against weight exfiltration. Potential mitigations include query rate limiting, output perturbation, and differential privacy techniques, though trade-offs with model performance remain unclear.
The findings underscore growing concerns about AI model security alongside recent debates over data privacy in machine learning development.
The Open Observatory of Network Interference (OONI) is expanding its crowdsourced effort to map global internet censorship. The project invites users to contribute measurements to what it describes as the largest open dataset on network interference.
A malicious npm campaign demonstrates how threat actors are evading supply chain protections by embedding malware in package runtime behavior instead of installation scripts. The 'indexed-btree' package exemplifies this evolving attack technique.
Cybercriminals are exploiting lookalike characters from different alphabets to create fake URLs that appear legitimate to the naked eye. These homoglyph attacks bypass traditional security checks and trick users into visiting malicious sites.
The ShinyHunters extortion gang has compromised the Clop ransomware operation's data leak site, defacing it and stealing server data and private encryption keys.