Published on: 15-Oct-2018
A research team led by Associate Professor Liu Yang (left) from the Cyber Security Lab received two ACM Distinguished Paper Awards for their papers “Deepguage: Multi-Granularity Testing Criteria for Deep Learning Systems” and “CIDiff: Generating Concise Linked Code Differences” at the 33rd IEEE/ACM International Conference on Automated Software Engineering, 3-7 September 2018, Montpellier, France.
Deep Learning (DL) has achieved tremendous success over the pass decades in many cutting edge applications, including safety-critical scenarios. However, a plethora of studies have shown that the state-of-the-art DL systems still suffer from various vulnerabilities which can lead to severe consequences when applied to real-world applications.
Their work DeepGauge proposes a comprehensive and multi-granular set of testing specially for DLs, which aims at rendering a multi-faceted portrayal and testing quality. This work builds the foundation of quality assurance and testing of DLs, and pioneers the new interdisciplinary research direction in software engineering, security and AI. The usefulness of DeepGauge has been demonstrated on state-of-the-art adversarial attacks as well as in consecutive follow-up research work. DeepGauge sheds light on the construction of more generic and robust DL systems, towards reshaping next generation intelligent software with safety and quality guarantees.
The second work, CIDiff, proposes a novel code differencing approach to generate a concise, linked representation of code differences, whose granularity is in between the existing code differencing and code change summarization methods. The proposed approach not only generates short and informative code differences, but also establishes their relationships. Experiments on 12 open-source Java projects (i.e. 74,387 commits) have demonstrated that the proposed approach generated more than 80% shorter code differences for 48% commits with 72% shorter time, and was more useful in change understanding than a state-of-the-art code differencing approach.
Back to listing