Beyond data poisoning in federated learning

The current boom of artificial intelligence (AI) is based upon neural networks (NNs). In order for these to be useful, the network has to undergo a machine learning (ML) process: work over a series of inputs, and adjust the inner weights of the connections between neurons so that each of the data samples the network was trained on produces the right set of labels for each item. Federated learning (FL) appeared as a reaction given the data centralization power that traditional ML provides: instead of centrally controlling the whole training data, various different actors analyze disjoint subsets of data, and provide only the results of this analysis, thus increasing privacy while analyzing a large dataset. Finally, given multiple actors are involved in FL, how hard is it for a hostile actor to provide data that will confuse the NN, instead of helping it reach better performance? This kind of attack is termed a poisoning attack, and is the main focus of this paper. The authors set out to research how effective can a hyperdimensional data poisoning attack (HDPA) be to confuse a NN and cause it to misclassify both the items trained on and yet unseen items.

Data used for NN training is usually represented as a large set of orthogonal vectors, each describing a different aspect of the item, allowing for very simple vector arithmetic operations. Thus, NN training is termed as high-dimensional or hyperdimensional. The attack method described by the authors employs cosine similarity, that is, in order to preserve similarity, a target hypervector is reflected over a given dimension, yielding a cosine-similar result that will trick ML models, even if using byzantine-robust defenses.

The paper is clear, though not an easy read. It explains in detail the mathematical operations, following several related although different threat models. The authors present the results of the experimental evaluation of their proposed model, comparing it to several other well-known adversarial attacks for visual recognition tasks, over pre-labeled datasets frequently used as training data, such as MNIST, Fashion-MNIST and CIFAR-10. They show that their method is not only more effective as an attack, but falls within the same time range as other surveyed attacks.

Adversarial attacks are, all in all, an important way to advance any field of knowledge; by publishing this attack, the authors will surely spark other works to detect and prevent this kind of alteration. It is important for AI implementers to understand the nature of this field and be aware of the risks that this work, as well as others cited in it, highlight: ML will train a computer system to recognize a dataset, warts and all; efficient as AI is, if noise is allowed into the training data (particularly adversarially generated noise), the trained model might present impaired performance.