Federated learning reduces the risk of information leakage, but remains vulnerable to attack. We show that well-mixed gradients provide numerical resistance to gradient inversion in neural networks. For example, we can enhance mixing gradients in a batch by choosing an appropriate loss function and drawing identical labels, and we support this with an approximate solution of batch inversion for linear layers. These simple architecture choices show no degradation to classification performance as opposed to noise perturbation defense. To accurately assess data recovery, we propose to use a variation distance metric for information leakage in images, derived from total variation. In contrast to Mean Squared Error or Structural Similarity Index metrics, it provides a continuous metric for information recovery. Finally, our empirical results of information recovery from various inversion attacks and training performance supports our defense strategies. These simple architecture choices found to be also useful for practical size of convolutional neural networks but depends on their size. We hope this work will trigger further defense studies using gradient mixing, towards achieving a trustful federation policy.

Distributed ledger technology benefits society by enabling an ecosystem of decentralised finance. However the pseudo-anonymised nature of transactions has also been an enabler of new routes for illicit activities ranging from individual scams to organised crimes. Current solutions for identifying addresses involved in illicit activities (illicit addresses) rely on commercial intelligence services, which are costly due to the intensive investigative efforts required. We propose Ledgit, an automatic real-time service for diagnosing illicit addresses on the Bitcoin blockchain. Ledgit is based solely on publicly available data, and uses an unsupervised clustering method that combines information from textual reports and the blockchain graph to assign a risk score that a Bitcoin address is involved in illicit activities. We verify the system with labeled addresses, showing high performance in identifying illicit addresses. Finally, we provide an intuitive user interface that provides accessible risk assessment with graph and report analytics.

Machine learning on source code (MLOnCode) is a popular research field that has been driven by the availability of large-scale code repositories and the development of powerful probabilistic and deep learning models for mining source code. Code-to-code recommendation is a task in MLOnCode that aims to recommend relevant, diverse and concise code snippets that usefully extend the code currently being written by a developer in their development environment (IDE). Code-to-code recommendation engines hold the promise of increasing developer productivity by reducing context switching from the IDE and increasing code-reuse. Existing code-to-code recommendation engines do not scale gracefully to large codebases, exhibiting a linear growth in query time as the code repository increases in size. In addition, existing code-to-code recommendation engines fail to account for the global statistics of code repositories in the ranking function, such as the distribution of code snippet lengths, leading to sub-optimal retrieval results. We address both of these weaknesses with Senatus, a new code-to-code recommendation engine. At the core of Senatus is De-Skew LSH a new locality sensitive hashing (LSH) algorithm that indexes the data for fast (sub-linear time) retrieval while also counteracting the skewness in the snippet length distribution using novel abstract syntax tree-based feature scoring and selection algorithms. We evaluate Senatus and find the recommendations to be of higher quality than competing baselines, while achieving faster search. For example on the CodeSearchNet dataset Senatus improves performance by 31.21% F1 and 147.9x faster query time compared to Facebook Aroma. Senatus also outperforms standard MinHash LSH by 29.2% F1 and 51.02x faster query time.

We show that knowledge of wallet addresses from the current time state of a blockchain network, such as Bitcoin, increases the performance of illicit activity detection. Based on this finding we introduce two new methods for the sampling of classifier training data so that precedence is given to transaction information from the recent past and the current time state. This sampling enables streaming classification in which a decision on the class of a transaction needs to be made based on data seen to date. Our new approach provides insight into how the dynamics of the blockchain network plays a central role in the detection of illicit transactions, and is independent of the classifier choice. Our proposed sampling methods enable graph convolution network (GCN) and random forest (RF) classifiers to better adapt to changes in the network due to significant events, such as the closure of a large ‘Darknet’ marketplace. We introduce Graphlet spectral correlation analysis for exposing the effect of such network re-organisation due to major events. Finally, based on our analysis, we propose a new two-stage random forest classifier that feeds back intermediate predictions of neighbours to improve the classification decision. Our methodology enables practical streaming classification, even in the scenario of very limited information on the feature space of each transaction.