Biblio

In this paper, we apply machine learning to distributed private data owned by multiple data owners, entities with access to non-overlapping training datasets. We use noisy, differentially-private gradients to minimize the fitness cost of the machine learning model using stochastic gradient descent. We quantify the quality of the trained model, using the fitness cost, as a function of privacy budget and size of the distributed datasets to capture the trade-off between privacy and utility in machine learning. This way, we can predict the outcome of collaboration among privacy-aware data owners prior to executing potentially computationally-expensive machine learning algorithms. Particularly, we show that the difference between the fitness of the trained machine learning model using differentially-private gradient queries and the fitness of the trained machine model in the absence of any privacy concerns is inversely proportional to the size of the training datasets squared and the privacy budget squared. We successfully validate the performance prediction with the actual performance of the proposed privacy-aware learning algorithms, applied to: financial datasets for determining interest rates of loans using regression; and detecting credit card frauds using support vector machines.

Distributed attacks originating from botnet-infected machines (bots) such as large-scale malware propagation campaigns orchestrated via spam emails can quickly affect other network infrastructures. As these attacks are made successful only by the fact that hundreds of infected machines engage in them collectively, their damage can be avoided if machines infected with a common botnet can be detected early rather than after an attack is launched. Prior studies have suggested that outgoing bot attacks are often preceded by other ``tell-tale'' malicious behaviour, such as communication with botnet controllers (C&C servers) that command botnets to carry out attacks. We postulate that observing similar behaviour occuring in a synchronised manner across multiple machines is an early indicator of a widespread infection of a single botnet, leading potentially to a large-scale, distributed attack. Intuitively, if we can detect such synchronised behaviour early enough on a few machines in the network, we can quickly contain the threat before an attack does any serious damage. In this work we present a measurement-driven analysis to validate this intuition. We empirically analyse the various stages of malicious behaviour that are observed in real botnet traffic, and carry out the first systematic study of the network behaviour that typically precedes outgoing bot attacks and is synchronised across multiple infected machines. We then implement as a proof-of-concept a set of analysers that monitor synchronisation in botnet communication to generate early infection and attack alerts. We show that with this approach, we can quickly detect nearly 80% of real-world spamming and port scanning attacks, and even demonstrate a novel capability of preventing these attacks altogether by predicting them before they are launched.