Characterization and Biological Activities of Essential Oils from Some Myrtaceae Species

Abstract

The pinnacle of networking is to build networks that coordinate, maintain, and run them selves. Such a dream seems achievable largely due to the advances in machine learning (ML), which has already transformed a variety of industries and impacted almost all areas of science. With networks becoming more heterogeneous, dense, and complex in nature, the application of ML for automating the networks is becoming popular. The last decade has witnessed an exponential increase in the ML-based solutions for automating different networking applications. To date, the networking community is skeptical in adapting the ML-based networking applications. This skepticism stems from many issues in the ML techniques such as lack of interpretability of ML techniques, lack of operational success, computational overhead, security vulnerabilities, lack of benchmarking dataset, etc. Network security in the presence of the adversary has always been a challenging problem. Recently, ML techniques have been shown to be brittle against small intelligently designed distortions in the input data introduced by the adversary for undermining the integrity of the ML model. These distortions are known as adversarial examples and the procedure of crafting and applying these techniques is known as adversarial attacks. Adversarial exam ples have become one of the major problems in the applicability of ML techniques in the wild. Since networking applications are security-critical, the adversarial attacks can have detrimental consequences on the performance of the network. We have worked on both adversarial attacks and defenses for ML-based networking applications. In this thesis, we have evaluated the performance of different ML-based networking appli cations (network traffic classification and anomaly-based intrusion detection system) against novel network-specific adversarial ML attacks. The goal of these attacks is to undermine the integrity of the ML-based networking applications. We have evaluated the performance of the proposed attacks on conventional and deep learning-based networking applications under different threat model assumptions and highlighted the lack of robustness in the ML-based networking applications. The robustness improvement against adversarial examples in the ML-based networking ap plications is another challenging problem. We have used both off-the-shelf and novel ro bustness improving procedures for defending against the adversarial examples in ML-based networking applications (network traffic classification and anomaly-based intrusion detec tion system). The preemptive defensive interventions proposed in this dissertation improves the performance of the ML-based networking applications in the presence of the adversary. Here we also want to note that defending against the adversarial attacks is still an open xviii problem and requires immediate attention. We have also highlighted how the game between adversaries (attacker and defender) is an arms race and what steps must be followed for evaluating any defensive intervention in ML-based networking applications. In summary, we identify and make an effort to bridge the research-gap between the increas ing approbation of ML techniques in networking and the lack of evaluation of the evolved threat of adversarial ML attacks. We also present a cautionary perspective on the use of ML in networking. This dissertation also calls for future investigation of adversarial ML attacks and defenses in networking while providing a comprehensive description of designing and application of adversarial attacks and defenses.