Research
158d agoIsolating Safety Features in Images
Researchers at Apple Machine Learning Research unveiled SafetyPairs, a method that isolates safety‑critical image features using counterfactual generation. By pinpointing the exact visual cues that shift an image from benign to harmful, the approach refines image safety datasets and enhances model interpretability.
The Signal
The technique, showcased at the ICLR 2026 workshop, promises more reliable AI systems worldwide.