CASPA 2019 Annual Conference & Dinner Banquet: AI from Silicon to Software

Fari Assad-Eraghi

SVP,NXP

Biography:

Fari Assad-Eraghi is Senior VP of Innovation and Advanced Technologies at NXP. He is responsible for technical roadmaps in all product and technology areas spanning computing, connectivity, AMS, and security. Prior to this role, he was CTO of IoT and Security, a Business Unit with $4B annual revenue. Among other responsibilities, he and his team set the roadmap and delivered security architecture and IP for NXP. Before joining NXP in 2016, he was VP of Advanced Technology at InvenSense. There he was tasked with expanding company’s product portfolio beyond inertial sensors to audio solutions, ultrasonic fingerprint sensors, and health sensors. Fari earned his M.S. and Ph.D. degrees in EECS from UC Berkeley. From 1995 to 2001 he was with IBM TJ Watson Research Center where he contributed to development of first commercial version of CMOS SOI for VLSI. At Rambus he and his teams developed high speed data communication circuits and systems, including the Terabyte/sec signaling initiative. As Senior VP of Engineering and Operations at SiTime, he led the first commercialization of MEMS-based timing products.

Fari has contributed to a wide technical area spanning semiconductor device physics, transducers, circuit design for analog and mixed-signal, low-power computing, and security. He has given numerous invited talks including at VLSI Symposium, CICC, RFIC, ISLPED, SOI, S3S and SDW Conferences. He is on IEEE’s North America and Europe Executive Committee. Dr. Assad-Eraghi has contributed to more than 100 technical papers, including the best paper at 2016 ISSCC, and has 70 issued patents.

Speech Title:Machine Learning and Trust

Abstract:

Machine Learning capabilities have been growing tremendously over the past 10 years due to advances in Deep Learning algorithms, acceleration of computational capabilities (both at cloud and edge), and availability of enormous amount of data. This has brought about many beneficial applications from speech recognition and machine translation to image and facial recognition. Countless other applications are being developed that will have a seminal impact on the entire human society. ML will transform all aspects of economy from healthcare (diagnostics, disease control, pharmaceuticals), to infrastructure (transport, logistics, energy), to finance, to advanced industries, and clearly to consumer. This realization has not only led to break-neck advances in computer science, software and algorithms, but also a renaissance in hardware innovations that address both processing and storage of data. Vast engineering resources have been directed towards improving performance and power efficiency of algorithms and their silicon counterparts. The efforts have been directed to both ends of the spectrum, from massive data centers consuming Mega-Watts of power and computing 100’s of petaflops, to end node IoT devices with milli-Watt power consumption and a few MIPS of compute resources. In this race for performance, until recently little attention was paid to the trust facet of Machine Learning. If any, the focus was only on improving accuracy of ML models. However, with the broad reach of ML finally being appreciated, many trust aspects are gaining attention and a burgeoning research field is taking shape. The Trust Umbrella covers security, privacy, interpretability, and fairness of ML. ML and security intersect in two ways: ML for security and security of ML. ML can be utilized to impact security of platforms and applications. It can enhance security as an additional defense mechanism by enabling tasks such as anomaly detection, intrusion detection, and control flow protection.

On the other hand, it can become a tool of bad actors by making exploitation of system vulnerabilities such as ‘side-channel attacks’ easier and more automated. Security of ML itself is however the larger concern. Both in the training phase and inferencing phase, researchers have shown that ML can be exploited. In training phase, data poisoning can lead to ML models that have specific vulnerabilities and ‘back doors’. In inferencing phase use of adversarial inputs can intentionally cause ML to mis-predict, compromising security and safety. Although recent research papers have demonstrated exploitation of ML vulnerabilities, and also countermeasures against these attacks, the practical nature and real-world applicability of these investigations are in question. Therefore, the threat models need to be much better defined and systematic methods for evaluating effectiveness of countermeasures developed.

Under this Trust Umbrella the other factor to consider is model confidentiality. Given the disproportionate amount of resources spent into training the models, and the enormous curated data that is essentially embedded in these models, model confidentiality is of prime importance. The motivation is not only avoidance of IP loss, but also the realization that attacks against ML become more efficient if adversaries have “white-box” access to the models. Another important factor is data confidentiality. Data used both in training phase and inferencing phase can be sensitive, particularly in Machine Learning as a Service (MLaaS) applications. Examples include health, financial, industrial, and governmental data. If private data is leaked, it can lead to regulatory compliance violations (e.g., GDPR), loss of customer trust, and damage to a company’s brand. Recently, several advances in privacy-preserving ML have been made that rely on cryptographic techniques. They span from fully-homomorphic encryption (FHE), to HE and garbled circuits and multi- party compute techniques. All these approaches have significant computational overhead and are mostly limited to inferencing phase at this point. The field is very nascent and rapid advances in hardware and algorithms are being made. Finally, a major disadvantage of current machine learning approaches is that insights about the data and the task the machine solves is hidden in increasingly complex models. If one focuses only on performance, the result is more and more opaque models. For example, the winning models of the recent Kaggle.com competition were mostly ensembles of models or very complex models such as boosted trees or deep neural networks. Interpretability is the degree to which a human can understand the cause of a ML decision. With more opaque and complex models, interpretability becomes extremely difficult. This is a major issue once we realize that ML can be used for many tasks where fairness and lack of bias are important such as credit rating, and policing, or where wrong decisions can have safety consequences. We are at the very early stages of evaluating ML models from this perspective and developing interpretable models.