Note: The order of speakers is sorted according to the alphabetical order of surname.
Dr. Tieyan Li
Head of AI Security, Shield Lab. Huawei.
On building full-stack AI security capabilities
AI is being widely adopted in industrial applications, while the evolution of AI attacks pose serious security threats. Building AI system may have security risk in every aspects including model, data, and software, e.g., including evasion attack, poisoning attack, backdoor attack, as well as model stealing attacks. The defense against these attacks has been hot research topics, proposed mitigations include input transformation, model transformation, and model testing and verification. In this talk, I will elaborate the AI threats and mitigations given the software and hardware AI stacks we are developing. I will briefly summarize the research challenges on evaluating the robustness of the DNN models and point out promising research directions ahead. Additionally, I will introduce a lightweight DNN-based malicious behavior detection scheme deployed in mobile phone. It achieves high accuracy on detecting dynamic behaviors and low power consumption with NPU acceleration.
Professor Chris Mitchell
Information Security Group, Royal Holloway, University of London
Quantum computing – What will be the real impact on security?
As has been widely discussed, if and when large-scale general-purpose quantum computers are constructed, the effect on currently used cryptography will be very significant – all asymmetric algorithms based on factoring large integers or discrete logarithms (include elliptic curve) will be rendered insecure for practically feasible key lengths. Also, all symmetric algorithms will in effect have their key length halved; i.e., with the aid of a quantum computer, a k-bit key could be ‘brute-forced’ in of the order of 2k/2 computations, so a 128-bit key will offer roughly the same level of security as a 64-bit key does today. In this context it is clearly vital to consider what the impact will be in practice. For algorithms solely used for verifying data integrity of transmitted data, there will no significant problems, as long as new `secure’ algorithms are introduced by the time quantum computing is available. However, the significance for encryption and key establishment algorithms is potentially catastrophic. If ciphertext is intercepted and stored, then it could be decrypted if and when the encryption algorithm is broken; i.e. data whose secrecy is of long-term significance is being made vulnerable right now through the use of algorithms which are quantum-unsafe.
These widely discussed facts mean that it is imperative that we carefully examine current uses of cryptography to understand what needs to be done and when. In this talk I will use a case study of the current 5G security provisions as the basis for a more general discussion of the quantum security problem. References will also be made to other widely used security schemes, such as SSL/TLS and the EMV payment security system. Conclusions will be drawn about what algorithms need changing and when, on the assumption that quantum computing will eventually become a reality.