Structured research. Real-world consequence.

Training data of a machine learning application may be collected from geographically different places, but transferring distributed data to a central server for training can be legally or practically impossible. This has led to a fast-growing area in machine learning, i.e., federated learning. However, it was discovered that the federated learning in its original form, where a server iteratively aggregates local gradients received from the distributed users, may suffer from privacy leakage as an attacker can infer sensitive information from the local gradients. Since then, many improvements have been presented to mitigate this privacy leakage, but they often achieve better privacy with the compromise of other critical properties such as the accuracy and communication complexity. In this paper, we propose a novel solution that is based on a fully decentralised distributed learning. Our solution enables the optimisation over, and achieve a balance between, multiple properties, including privacy preservation, accuracy, communication complexity, efficiency, and tolerance to user failures. Our solution proceeds by first synthesising a communication topology between users according to the required properties and then applying a fully decentralised distributed learning where the server is not involved in the computation. In the decentralised learning, the aggregation of local gradients is reduced to a distributed consensus between users. Finally, the agreed value of the users is sent to the server after added a differential privacy noise. We are conducting experiments on both use cases in the US/UK privacy enhancing technologies challenge to validate our solution.

Ubiquitous AI will soon allow complex systems to drive on our roads, fly over our heads, move alongside us during our daily lives & work in our factories. In spite of this disruptive landscape, deployment and broader adoption of learned-enabled autonomous systems in safety-critical scenarios remains challenging. Continuous engineering (DevOps) can mediate problems when encountering new scenarios throughout the product life cycle. However, the technical foundations and assumptions on which traditional safety engineering principles rely do not extend to learning-enabled autonomous systems engineered under continuous development. FOCETA gathers prominent academic groups & leading industrial partners to develop foundations for continuous engineering of trustworthy learning-enabled autonomous systems. The targeted scientific breakthrough lies within the convergence of “data-driven” and “model-based” engineering, where this convergence is further complicated by the need to apply verification and validation incrementally & avoid complete re-verification & re-validation efforts. FOCETA’s paradigm is built on three scientific pillars: (1) integration of learning-enabled components & model-based components via a contract-based methodology which allows incremental modification of systems including threat models for cyber-security, (2) adaptation of verification techniques applied during model-driven design to learning components in order to enable unbiased decision making, & finally, (3) incremental synthesis techniques unifying both the enforcement of safety & security-critical properties as well as the optimization of performance. FOCETA approach, implemented in open source tools & with open data exchange standards, will be applied to the most demanding & challenging applications such as urban driving automation & intelligent medical devices, to demonstrate its viability, scalability & robustness, while addressing European industry cutting-edge technology needs.

EnnCore will deliver safeguarding for safety-critical neural-based architectures with the following novel properties: (1) Full-stack symbolic software verification: we will develop the first bit-precise and scalable symbolic verification framework to reason over actual implementations of DNNs, thereby providing further guarantees of security properties concerning the underlying hardware and software, which are routinely ignored in existing literature. (2) Explainability / Interpretability: EnnCore will pioneer the integration of knowledge-based and neural explainability methods to support end-users specifying security constraints and diagnosing security risks, in order to provide assurances about its security as NB models evolve. Particular attention will be given to the quantitative and qualitative characterization of semantic-drift phenomena in security scenarios. (3) Scalable: we will systematically combine contemporary symbolic methods for explaining, interpreting and verifying neural representations. In particular, we will develop a neuro-symbolic safeguard framework by linking the structural knowledge-based representation elements to the attentional architecture elements, to achieve scalability and precision in an unprecedented manner. We will also develop new learning techniques for reusing information across different verification runs to reduce formulae size and consistently to improve constraint solving.