Safexplain DIVRED: Software-only image transformation-based AI diverse redundancy

Software-only image transformation-based AI diverse redundancy

Institution:

Technology

Research Group:

BSC Group: High-Performance Embedded Systems Lab

Researcher/s:

Martí Caro Roca, Axel Brando Guillaumes, Jaume Abella Ferrer, Jordi Fornt Mas

Description:

SafeExplain-DIVRED improves the reliability of AI vision systems by testing multiple equivalent slight variations of the same input image and combining results. Smart combination approaches allow filtering out problematic results due to model limitations, random errors, or even adversarial attacks. This approach adds a safety layer without modifying the original model, supporting trustworthy AI. It can be applied to any object detection model and can be ported to virtually any platform.

Value Proposition:

Add run-time safety wrapper to any AI model-no retrain

Aplication areas:

Autonomous-vehicle perception & planning, surgical-robot vision, industrial defect inspection, UAV sense-and-avoid, legal/medical LLM assistants, critical infrastructure anomaly detection.

Target market:

Automative Tier-1s & robot OEMs deploying vision/NLP models; Medical-device manufacturers under IEC 62304 & MDR; Aerospace UAV and satellite autonomy stacks; Cloud & edge providers offering "regulated AI" services

Technology Readiness Level (1-9): N/A

Protection:

Proprietary software

More information

if you want to know more about this project do not hesitate to contact us

Contact us