Secure Machine Learning & Sensor Instrumentation in a Trusted Architecture

1. Secure Machine Learning
a. Digital twins are becoming more prevalent in industry, comprising of a series of connected models emulating a physical process. Machine learning and AI form much of the data driven decisions and many common techniques employ 'black box' techniques in which you can't really inspect how the algorithm is accomplishing what it is accomplishing. To develop trust in an algorithm you must have trust in the input and output data, both of which can be subject to interference directly or indirectly. Therefore several questions arise,
i. How can you trust that the output you receive from such a technique has not been tampered with in some way?
ii. Are you training your classifier on the correct data?
iii. Has the input data been manipulated?
iv. Has somebody skewed a test case to influence your classifier?
v. How can you quantify the quality/safety of AI processes?
2. Secure Sensor Instrumentation & Control systems
a. Pharmaceutical manufacturing contains highly sensitive processes, dependent on numerous inputs both in terms of materials and data. However, these are vulnerable to attack either through malicious intent, inaccuracies in the supply chain or incorrect machine operation. The impact of inaccuracies in the pharmaceutical processes could be life threatening to a large number of people, as well as extremely damaging to the reputation of a company/medicine. Control operations for continuous manufacturing must ensure the safety of a particular operation, however how can we ensure that this process has not been tampered with in any way?
i. Can we combine control systems with secure machine learning for added security?
ii. How do we verify the traceability of materials which enter the control process?
iii. Can we ensure attack resistant continuous manufacturing?