Adversarial Program Synthesis

Malware classifiers are strongly affected by concept drift: the statistical properties of new malware examples change with respect to those the classifier was trained with and the lack of i.i.d. induces decay in the performance of ML tasks over time. This has resulted in an endemic issue that permeates the security community: in these contexts, ML-driven techniques lack an unbiased evaluation in realistic settings. In malware classification and detection tasks representative of such a problem---when the i.i.d. assumption of training and testing datasets do not hold anymore---traditional ML evaluation methodologies such as k-fold cross-validation produce inflated results due to time and space inconsistencies.

Additionally, classifiers are susceptible to adversarial attacks which can be characterized as an extreme form of concept drift. Here the classifier is induced into making incorrect decisions by an attacker that alters data at train-time (poisoning) or at test-time (evasion). In malware detection, this implies that an adversarial malware object is erroneously recognized as goodware by the classifier to bypass antivirus solutions. Most research efforts on adversarial ML have focused on the image domain, where it is trivial to change pixels; the malware domain is much more challenging because program code must be altered while preserving functionality. Hence, there is no clear understanding of adversarial attacks in the malware domain. This lack of understanding hinders the design, development and evaluation of robust machine-learning based defences against future malware techniques.

This project aims to study which adversarial attacks for malware are realistic (i.e. which adversarially-generated feature vectors can be transformed into real and functioning applications), how to employ novel program analysis and machine learning techniques to defend against such threats, and how to evaluate complex defensive systems in a fair and systematic way.

The most significant impact of the renewal of Royal Holloway's CDT in Cyber Security will be the production of at least 30 further PhD-level graduates. In view of the strong industry involvement in both the taught and research elements of the programme, CDT graduates are "industry-ready": through industry placements, they have exposure to real-world cyber security problems and working environments; because of the breadth of our taught programme, they gain exposure to cyber security in all its forms; through involvement of our industrial partners at all stages of the programme, the students are regularly exposed to the language and culture of industry. At the same time, they will continue to benefit from generic skills training, equipping them with a broad set of skills that will be of use in their subsequent workplaces (whether in academia, industry or government). They will also engage in PhD-level research projects that will lead to them developing deep topic-specific knowledge as well as general analytical skills.

One of the longer-term impacts of CDT research, expressed directly through research outputs, is to provide mechanisms that help to enhance confidence and trust in the on-line society for ordinary citizens, leading in turn to quality of life enhancement. CDT research has the potential of directly impacting the security of deployed system, for example helping to make the Internet a more secure place to do business. Moreover the work on the socio-technical dimensions of security and privacy also gives us the means to influence government policy to the betterment of society at large. Through the training component of the CDT, and subsequent engagement with industry, our PhD students are exposed to the widest set of cyber security issues and forced to think beyond the technical boundaries of their research. In this way, our CDT is training a generation of cyber security researchers who are equipped - philosophically as well as technically - to cope with whatever cyber security threats the future may bring. The programme equip students with skills that will enable them to understand, represent and solve complex engineering questions, skills that will have an impact in UK industry and academic long beyond the lifetime of the CDT.