Tag
Apple introduces 'Apple Reference Image,' a secure camera mode for iPhone that ensures verifiable photography with privacy protection, debuting on iPhone 18 Pro and Pro Max models.
The article introduces the world's first double-blind evaluation of a proprietary AI model, using cryptographic environments to prevent benchmark contamination and enhance trust in AI safety assessments.
This paper explores machine unlearning techniques for Large Audio-Language Models to remove sensitive information from speech QA tasks, demonstrating methods that reduce privacy leakage by up to 80% while maintaining performance.
The paper argues that federated learning is not a complete remedy for extractive AI, focusing on governance layers for creative communities and proposing design principles for a data commons that governs models and federation.
The paper introduces a transparent framework that maps acoustic speech features to DSM-5 depression indicators for interpretable detection, running locally on commodity hardware to preserve privacy.
Google DeepMind is piloting the world's first double-blind evaluations for frontier AI models to ensure secure and trustworthy external assessments, partnering with organizations like Singapore AI Safety Institute and MLCommons.
Anthropic has enabled external researchers to independently study AI impacts using real, privacy-preserved Claude usage data through its Anthropic Insights tool, marking the first time such data has been made accessible outside AI labs.
PANDA is a scalable system using zero-knowledge proofs to verify the robustness and fairness of neural networks without revealing model parameters, enabling certification for large networks with polynomial complexity.
CutClean is a privacy-aware pruning method for neural networks that reduces private information leakage while increasing sparsity, using auxiliary privacy heads to quantify and mitigate privacy risks.
The article compares confidential inference APIs from Privatemode, Tinfoil, NEAR AI, and Chutes, highlighting their security features like end-to-end encryption and trusted execution environments, along with tradeoffs in model selection and verification maturity.
QFedPolyp proposes a federated learning framework for polyp segmentation that uses quantization-aware training to reduce communication costs and achieve faster inference while preserving privacy.
This paper introduces Repeated Deceptive Path Planning (RDPP) and a novel framework called DeceptiveMetaPlanning (DeMP) to enable agents to maintain deception against observers that learn and adapt over time.