Elevating trust on AI by software Sample Clauses

Elevating trust on AI by software security‌ To define how and where software security is useful for elevating AI/ML security, a good start is a focused technical survey categorizing and detailing the modes of attack and the classical existing mitigations. Before that, however, it is appropriate to state the discrepancies between a general- purpose data processing algorithm and AI/ML ones, in the perspective of their security. Our survey is produced at two perimeters successively. First, we consider any type of AI/ML application. Secondly, we look more precisely at the AI/ML systems implemented on 5G networks. This scope restriction is aimed at establishing what are the most salient threats in AI/ML in operation on 5G networks. We can better elaborate the attack areas laid over the data extraction and processing pipeline as well as assess if software security can bring any valuable contribution there. When this work is completed, we produce a specific technical survey on the merits of trusted execution environment to frustrate the model stealing. We then conclude by providing the results and lessons learnt from our technical investigation of Intel SGX used as a shield for the model as well as other techniques applied on the data collection and inspired from this research work. 4.2.1 Specificities of AI/ML solutions (in view of their software security)‌ AI/ML systems are not developed and conceived by the application developer as are (any) other data processing software. The divergence relies on the core code (learning algorithm) genesis, not developed by the application developers but conversely by machine learning framework teams. Algorithm and framework are common bricks used for several very different applications by the AI/ML (vertical) solution designers. Some depicts this situation as “AI is automatic programming” which is a factual statement as no one needs to program it. A common pre-existing software is customized and tuned automatically at the training stage resulting in the production of the model parameters. The added value or intellectual property of the AI/ML vertical application designer is in the appropriate selection of the (pre-existing) algorithm and its training with the well- or best-fitted training data. A stock phrase is that AI/ML quality depends on its training data. As a consequence, with AI/ML, the intellectual property resides precisely on the model’s parameters which consists in a data structure, not an algorithm. This entails one critical Intellectual Property Ri...