Detecting and fending off attacks on computer systems is an enduring problem in computer security. In light of a plethora of different threats and the growing automation used by attackers, we are in urgent need of more advanced methods for attack detection. Manually crafting detection rules is by no means feasible at scale, and automatically generated signatures often lack context, such that they fall short in detecting slight variations of known threats.
In the thesis “Efficient Machine Learning for Attack Detection” , we address the necessity of advanced attack detection. For the effective application of machine learning in this domain, a periodic retraining over time is crucial. We show that with the right data representation, efficient algorithms for mining substring statistics, and implementations based on probabilistic data structures, training the underlying model for establishing an higher degree of automation for defenses can be achieved in linear time.
External Human-Machine Interfaces (eHMIs) are expected to bridge the communication gap between an automated vehicle (AV) and pedestrians to replace the missing driver-pedestrian interaction. However, the relative impact of movement-based implicit communication and explicit communication with the aid of eHMIs on pedestrians has not been studied and empirically evaluated. In this study, we pit messages from an eHMI against different driving behaviors of an AV that yields to a pedestrian to understand whether pedestrians tend to pay more attention to the motion dynamics of the car or the eHMI in making road-crossing decisions. Our contributions are twofold: we investigate (1) whether the presence of eHMIs has any objective effect on pedestrians’ understanding of the vehicle’s intent, and (2) how the movement dynamics of the vehicle affect the perception of the vehicle intent and interact with the impact of an eHMI. Results show that (1) eHMIs help in convincing pedestrians of the vehicle’s yielding intention, particularly when the speed of the vehicle is slow enough to not be an obvious threat, but still fast enough to raise a doubt about a vehicle’s stopping intention, and (2) pedestrians do not blindly trust the eHMI: when the eHMI message and the vehicle’s movement pattern contradict, pedestrians fall back to movement-based cues. Our results imply that when explicit communication (eHMI) and implicit communication (motion-dynamics and kinematics) are in alignment and work in tandem, communication of the AV’s yielding intention can be facilitated most effectively. This insight can be useful in designing the optimal interaction between AVs and pedestrians from a user-centered design perspective when driver-centric communication is not available.
Secure connections are at the heart of today’s Internet infrastructure, protecting the confidentiality, authenticity, and integrity of communication. Achieving these security goals is the responsibility of cryptographic schemes, more specifically two main building blocks of secure connections. First, a key exchange protocol is run to establish a shared secret key between two parties over a, potentially, insecure connection. Then, a secure channel protocol uses that shared key to securely transport the actual data to be exchanged. While security notions for classical designs of these components are well-established, recently developed and standardized major Internet security protocols like Google’s QUIC protocol and the Transport Layer Security (TLS) protocol version 1.3 introduce novel features for which supporting security theory is lacking.
In my dissertation , which this article summarizes, I studied these novel and advanced design aspects, introducing enhanced security models and analyzing the security of deployed protocols. For key exchange protocols, my thesis introduces a new model for multi-stage key exchange to capture that recent designs for secure connections establish several cryptographic keys for various purposes and with differing levels of security. It further introduces a formalism for key confirmation, reflecting a long-established practical design criteria which however was lacking a comprehensive formal treatment so far. For secure channels, my thesis captures the cryptographic subtleties of streaming data transmission through a revised security model and approaches novel concepts to frequently update key material for enhanced security through a multi-key channel notion. These models are then applied to study (and confirm) the security of the QUIC and TLS 1.3 protocol designs.
In this paper we present a methodology to automatically generate an accurate behavioral model from an analog circuit description. The current machine learning method is limited to circuits with up to 80 transistors, limiting our approach to small and mid size circuit blocks due to a state explosion problem. However, if complex building blocks such as IOT systems should be modeled, the current approach needs to recoup with feasible simulation and modeling time. To come up with a solution for this problem, we extend the current method by a compositional approach. The approach is illustrated upon an example from the area of autonomous driving. Our method decomposes this large example into smaller building blocks and models each of them automatically. All models are combined into a compositional hybrid automaton of the whole complex system.
Compared to the original state space, the building blocks operate on smaller and reduced state spaces and hence drastically reduce the complexity. Using a back-transformation on the compositional automaton, all values from the original state space can be reconstructed. Moreover, we perform a formal verification on the generated compositional automaton. Results from a meaningful example are presented and discussed.