Model theft is the unauthorized replication or extraction of an AI model’s capabilities. Attackers repeatedly query an API, collect outputs, and use them to train a substitute model, avoiding the original development and training costs. section 11.1.3
Recommendation API cloning
An attacker could repeatedly query DataMax’s recommendation API and use the responses to reverse-engineer a competing clone. This could weaken DataMax’s competitive advantage, enable cheaper imitation, and potentially expose sensitive training data. section 11.1.3
Reduce model-theft risk with layered technical, contractual, and privacy protections:
- Limit API rates to restrict extensive querying and data extraction. section 11.1.3
- Apply watermarking to outputs so unauthorized use can be detected. section 11.1.3
- Use differential privacy to protect individual data points if the model is compromised. section 11.1.3
- Use homomorphic encryption to compute on encrypted data without decrypting it first. section 11.1.3
- Enforce end-user license agreements and intellectual-property rights against unauthorized replication. section 11.1.3
