Abstract-Several studies have shown that the network traffic that is generated by a visit to a website over Tor reveals information specific to the website through the timing and sizes of network packets. By capturing traffic traces between users and their Tor entry guard, a network eavesdropper can leverage this meta-data to reveal which website Tor users are visiting. The success of such attacks heavily depends on the particular set of traffic features that are used to construct the fingerprint. Typically, these features are manually engineered and, as such, any change introduced to the Tor network can render these carefully constructed features ineffective. In this paper, we show that an adversary can automate the feature engineering process, and thus automatically deanonymize Tor traffic by applying our novel method based on deep learning. We collect a dataset comprised of more than three million network traces, which is the largest dataset of web traffic ever used for website fingerprinting, and find that the performance achieved by our deep learning approaches is comparable to known methods which include various research efforts spanning over multiple years. The obtained success rate exceeds 96% for a closed world of 100 websites and 94% for our biggest closed world of 900 classes. In our open world evaluation, the most performant deep learning model is 2% more accurate than the state-ofthe-art attack. Furthermore, we show that the implicit features automatically learned by our approach are far more resilient to dynamic changes of web content over time. We conclude that the ability to automatically construct the most relevant traffic features and perform accurate traffic recognition makes our deep learning based approach an efficient, flexible and robust technique for website fingerprinting.

The adoption of machine learning and deep learning is on the rise in the cybersecurity domain where these AI methods help strengthen traditional system monitoring and threat detection solutions. However, adversaries too are becoming more effective in concealing malicious behavior amongst large amounts of benign behavior data. To address the increasing time-to-detection of these stealthy attacks, interconnected and federated learning systems can improve the detection of malicious behavior by joining forces and pooling together monitoring data. The major challenge that we address in this work is that in a federated learning setup, an adversary has many more opportunities to poison one of the local machine learning models with malicious training samples, thereby influencing the outcome of the federated learning and evading detection. We present a solution where contributing parties in federated learning can be held accountable and have their model updates audited. We describe a permissioned blockchain-based federated learning method where incremental updates to an anomaly detection machine learning model are chained together on the distributed ledger. By integrating federated learning with blockchain technology, our solution supports the auditing of machine learning models without the necessity to centralize the training data. Experiments with a realistic intrusion detection use case and an autoencoder for anomaly detection illustrate that the increased complexity caused by blockchain technology has a limited performance impact on the federated learning, varying between 5 and 15%, while providing full transparency over the distributed training process of the neural network. Furthermore, our blockchain-based federated learning solution can be generalized and applied to more sophisticated neural network architectures and other use cases.

Numerous research studies have demonstrated the effectiveness of machine learning techniques in application to network intrusion detection. And yet, the adoption of machine learning for securing large-scale network environments remains limited. The community acknowledges that network security presents unique challenges for machine learning, and the lack of training data representative of modern traffic remains one of the most intractable issues. New attempts are continuously made to develop high quality benchmark datasets and proper data collection methodologies. The CICIDS2017 dataset is one of the recent results, created to meet the demanding criterion of representativeness for network intrusion detection. In this paper, we revisit CICIDS2017 and its data collection pipeline and analyze correctness, relevance and overall utility of the dataset for the learning task. In the process, we uncover a series of traffic generation, flow extraction and labeling problems that severely affect the validity of data. To counter these issues, we investigate and eliminate their causes and regenerate the dataset. As a result, more than 20% of traffic traces are reconstructed or relabeled, and machine learning benchmarks demonstrate significant improvements. Our study exemplifies how data collection issues may have enormous impact on model evaluation and provides recommendations for their anticipation and prevention.

With rapidly evolving threat landscape surrounding malware, intelligent defenses based on machine learning are paramount. In this chapter, we review the literature proposed in the past decade and identify the state-of-the-art in various related research directionsmalware detection, malware analysis, adversarial malware, and malware author attribution. We discuss challenges that emerge when machine learning is applied to malware. We also identify the key issues that need to be addressed by the research community in order to further deepen and systematize research in the malware domain.