Perfectly parallel fairness certification of neural networks

Abstract

Recently, there is growing concern that machine-learned software, which currently assists or even automates decision making, reproduces, and in the worst case reinforces, bias present in the training data. The development of tools and techniques for certifying fairness of this software or describing its biases is, therefore, critical. In this paper, we propose a perfectly parallel static analysis for certifying fairness of feed-forward neural networks used for classification of tabular data. When certification succeeds, our approach provides definite guarantees, otherwise, it describes and quantifies the biased input space regions. We design the analysis to be sound, in practice also exact, and configurable in terms of scalability and precision, thereby enabling pay-as-you-go certification. We implement our approach in an open-source tool called libra and demonstrate its effectiveness on neural networks trained on popular datasets. CCS Concepts: • Software and its engineering → Formal software verification; • Theory of computation → Program analysis; Abstraction; • Computing methodologies → Neural networks; • Social and professional topics → Computing / technology policy.

Reduced Products of Abstract Domains for Fairness Certification of Neural Networks

“…In our experimental evaluation we evaluate Libra on neural networks trained on a popular dataset and we demonstrate its effectiveness. In particular, we show that Libra (configured to use the product domain) outperforms its preliminary version [25] in terms of both precision and running time.…”

“…Table 2 in Section 3). Ultimately however, the optimal configuration largely depends on the analyzed neural network [25]. For this reason, we have equipped Libra with a configuration auto-tuning mechanism, which dynamically updates the lower bound and upper bound configuration according to a chosen search heuristic.…”

“…In the preliminary version of Libra [25], feasible partitions are first grouped by activation pattern, i.e., activation patterns that fix more ReLUs are merged with those that fix fewer ReLUs. This way, in principle, the amount of work that the backward analysis has to do is reduced: it only needs to run once for each activation pattern, and can then perform all the checks for bias on each feasible partition.…”

“…The pre-analysis can be configured to use any of the abstract domains implementations that Libra is equipped with. A preliminary version of Libra developed by Urban et al [25] was equipped with the boxes [4], symbolic [18], and the deeppoly [24] abstract domains. In our tool, we additionally implemented the neurify [26] abstract domain, and a generic reduced product domain construction [6] to combine any of these domains together.…”

Automated Safety Verification of Programs Invoking Neural Networks

“…Verily builds on Marabou [38], a verification tool for neural networks, and aims to ensure that a system achieves desired service-level objectives (expressed as safety or liveness properties). Other techniques use abstract interpretation to verify robustness [27,57,41] or fairness properties [63] of neural networks. Furthermore, there are several existing techniques for check-ing properties of neural networks using SMT solvers [37,38,36] and global optimization techniques [54].…”