Pangram Labs has launched Pangram Image 1.0, an AI-generated image detector that uses heatmaps to localize synthetic regions, alongside the Image Flash 1.0 video detector. The company reports 99.96% accuracy on its own 1,200-image benchmark and 96.4% at 256x256 resolution, though the figures have not yet been independently verified.
Termsslug: adversarial-robustness
Adversarial Robustness
对抗鲁棒性
The ability of a model to maintain stable predictions when facing maliciously constructed input perturbations.
Curated Intelligence
IT之家 · 智能时代T2·Models
58 pts
OpenAI 官方博客T1·Models
35 pts
This research proposes enhancing adversarial robustness by increasing inference-time compute. The core idea shifts defense burden from training to inference, leveraging extra computation for verification or correction. Details on methods, experiments, and performance trade-offs are currently missing.
OpenAI 官方博客T1·Models
25 pts
The paper investigates the transferability of adversarial robustness across different perturbation types. Due to the lack of body text, specific methods and results cannot be evaluated.