I am Professor of Computer Science at the University of Liverpool. Prior to Liverpool, I worked at the University of Oxford and the University of New South Wales in Australia. I am the Head of Artificial Intelligence for the School of Computer Science and Informatics, and prior to this, I was on the role of school research lead of the School of Electronic Engineering, Electronics, and Computer Science.

I founded the Trustworthy Autonomous Cyber-Physical Systems (TACPS) Laboratory, now located at the Digital Innovation Facility (DIF). Our mission: every AI system deployed in a safety-critical setting should come with a certification guarantee.

Research Interests

I work on making AI systems certifiably safe — developing the theory, algorithms, and tools needed so that neural networks, large language models, and autonomous agents can be deployed in safety-critical settings with provable guarantees. My group has been at the forefront of this field since publishing one of the first DNN verification algorithms (DLV, CAV 2017), and our work now spans the full certification stack from formal verification through to runtime assurance. I am the author of the textbook Machine Learning Safety (Springer, 2023; 61,000+ accesses as of April 2026), coordinator of the €9.3M EU Horizon project RobustifAI, and have co-chaired the AISafety@IJCAI and SafeAI@AAAI workshop series every year since 2019 — among the longest-running dedicated AI safety workshops at top AI conferences. My publications have received over 8,800 citations as of April 2026 (Google Scholar).

Our research programme has progressed through four paradigms, each addressing limitations of the previous one:

Earlier work on logic-based reasoning — strategic logics, model checking, and epistemic reasoning in multi-agent systems — provides the formal foundations that underpin our specification languages for learning-enabled components.

A walk-through of our research programme is available in these research overview slides.

I co-organise the Turing interest group on Neuro-Symbolic AI and contribute to standards bodies including SAE G-34/EUROCAE WG-114 (aviation AI). Our group alumni now hold faculty and research positions at Exeter, Imperial College London, Southampton, Warwick, Manchester, MBZUAI, and other institutions worldwide. I am a senior member of IEEE.

The research has been funded by EPSRC, European Commission, Dstl, Innovate UK, the Alan Turing Institute, the UK AI Security Institute, etc. I have been the PI (or Liverpool PI) for projects valued more than £10M, and co-I for more than £20M. Some brief information can be found here.

I led a team that won the UK-US privacy-enhancing technologies prize challenges at the first stage and a special recognition prize on "Novel Modelling/Design" at the second stage.

Machine Learning Safety book cover
Book "Machine Learning Safety" has been published by Springer.

Major Ongoing Projects

RobustifAI — €9.3M EU Horizon Europe · 18 Partners · 11 Countries · 2025–2028

Robustifying Generative AI through Human-Centric Integration of Neural and Symbolic Methods. Coordinated by Liverpool, with industry partners Collins Aerospace, Siemens, and Thales, and universities including Hebrew University (Katz), TU Wien, Chalmers, and others.

Three innovation axes: Neural-Symbolic Methods (combining logic/formal verification with ML for next-generation GenAI), Adaptiveness (introspection for hallucination detection and distribution shift), and Human Centricity (diverse stakeholders engaged throughout the lifecycle). Three use cases: autonomous driving (operational robustness), service robotics for patient care (user robustness), and cybersecurity SOC (technical robustness).

The Alignment Project

Funded through the Alignment Project, led by the UK AI Security Institute with EPSRC as a coalition partner, focusing on the development of rare-event estimation algorithms with targeted applications on AI agents with respect to jailbreaks and social deceptive behaviours.

Our research has spanned all stages of a full-stack safety case — testing, verification, safety assurance, and runtime monitoring. These have been validated in the SOLITUDE project (Dstl, 2020–2022) on an autonomous underwater vehicle, covering both simulated and physical executions in a controlled environment.

Open-Source Tools

We develop and maintain widely used safety verification and testing tools, including DLV (one of the first SMT-based DNN verifiers), DeepConcolic (concolic testing for DNNs), TrustAI (a suite of robustness analysis tools), and testRNN (coverage-guided testing for recurrent networks).

The following are a few video demos of our research:

For Prospective Students

I am always looking for PhD students with strong motivation to actively participate in research. There are a few possible ways of receiving a scholarship, for example:

If you have other means of supporting your study, you are also welcomed to get in touch.

New Open Positions

We regularly have openings for postdocs and PhD students. Please check back or get in touch if you are interested.

Workshop Organisation

Past workshops (2018–2022) ▸

Recent News

Older news (2020–2023) ▸
  • (12/2023) Three papers were accepted to AAAI-24. Congratulations to Zihao, Ronghui, Sihao, and all other co-authors.
  • (10/2023) We won an Alan Turing project "CRoCS: Certified Robust and Scalable Autonomous Operation in Cyber Space".
  • (08/2023) Our paper "Hierarchical Distribution-Aware Testing of Deep Learning" is accepted by ACM Transactions on Software Engineering and Methodology.
  • (07/2023) One paper accepted to ACM MM 2023.
  • (07/2023) Paper "SAFARI: Versatile and Efficient Evaluations for Robustness of Interpretability" accepted to ICCV2023.
  • (02/2023) Paper "Randomized Adversarial Training via Taylor Expansion" accepted to CVPR2023.
  • (01/2023) Paper "Decentralised and Cooperative Control of Multi-Robot Systems through Distributed Optimisation" accepted to AAMAS2023.
  • (12/2022) Textbook "Machine Learning Safety" published by Springer.
  • (11/2022) Paper "Towards Verifying the Geometric Robustness of Large-scale Neural Networks" accepted to AAAI2023.
  • (10/2022) Start co-organising Turing interest group on Neuro-symbolic AI.
  • (10/2022) Awarded a project on UK and US governments launched challenge on privacy-enhancing technologies (PETs).
  • (07/2022) Paper "Adversarial Label Poisoning Attack on Graph Neural Networks via Label Propagation" accepted to ECCV2022.
  • (06/2022) Two papers accepted to IROS 2022.
  • (03/2022) Paper on "enhancing adversarial training with second order statistics of weights" accepted to CVPR2022.
  • (10/2021) Congratulations to Yanda, who has three papers published at ICCV2021, IEEE Trans. Medical Imaging, and MICCAI2021.
  • (08/2021) Delivered a tutorial to IJCAI'2021 on "Towards Robust Deep Learning Models: Verification, Falsification, and Rectification".
  • (07/2021) One paper accepted by ICCV2021.
  • (07/2021) Paper "Embedding and Synthesis of Knowledge in Tree Ensemble Classifiers" accepted by Machine Learning journal.
  • (05/2021) Paper "BayLIME: Bayesian Local Interpretable Model-Agnostic Explanations" accepted to UAI2021.
  • (05/2021) Paper "Coverage Guided Testing for Recurrent Neural Networks" accepted to IEEE Trans. Reliability.
  • (10/2020) Started a new project "SOLITUDE: Safety Argument for Learning-enabled Autonomous Underwater Vehicles."
  • (09/2020) Paper "How does Weight Correlation Affect Generalisation Ability of Deep Neural Networks?" accepted to NeurIPS2020.
  • (08/2020) Paper "PRODEEP: a platform for robustness verification of deep neural networks" accepted to ESEC/FSE2020.
  • (07/2020) Two papers accepted to ECCV2020.
  • (06/2020) Paper "Practical Verification of Neural Network Enabled State Estimation System for Robotics" accepted to IROS2020.
  • (05/2020) Survey paper "A Survey of Safety and Trustworthiness of Deep Neural Networks" accepted to Computer Science Survey.

Teaching for this semester

Below are the research directions we have fostered over the past years. We use [Journal Name, Year] to denote a journal publication and [ConferenceAbbreviation+Year] to denote a conference paper.

Research Programme: From Formal Verification to GenAI Safety

Over the past decade, our work has progressed through four paradigms — each building on the limitations discovered in the previous one — in pursuit of suitable yet provable guarantees for AI safety.

Act I

Formal Verification

SMT-based DNN verification (DLV, CAV 2017). The gold standard — but NP-hard and doesn't scale to production networks.

Act II

Tightening Bounds

PAC-Bayesian bounds via weight correlation, spectral norms, and Taylor expansion. Tighter, but still loose for large models.

Act III

Statistical Guarantees

Conformal prediction and randomised smoothing — distribution-free, per-input certificates that scale. The practical sweet spot.

Act IV

GenAI Safety

Extending certification to LLMs, VLMs, diffusion models, and AI agents. Rare-event estimation, guardrails, and layered safety.

See the full narrative in our research overview slides.

The detailed research directions below map onto this progression:

(a) Surveys, Textbook, and Foundations

References: [Information Fusion, 2026], [Computer Science Review, 2026], [Artificial Intelligence Review, 2025], [Artificial Intelligence Review, 2024], [JLAMP, 2024], [Machine Learning Safety, Springer 2023], [ICFEM2022], [AI Communications, 2022], [CIKM2021], [Robotics, 2021], [Computer Survey Review, 2020]

(b) Formal Verification of Neural Networks and Learning-Enabled Systems

References: [ICSE2025], [AAAI-2025b], [AAAI-2025a], [ECCV2024-a], [Neurocomputing, 2024], [RA-L, 2024], [AAAI2023], [IEEE RA-L, 2023b], [PAKDD2023], [IROS2022a], [Formal Aspect of Computing, 2021], [ICANN2021], [Theoretical Computer Sciences, 2020], [FSE2020], [IROS2020], [IJCAI2019], [SAS2019], [TACAS2018], [IJCAI2018a], [CAV2017]

(c) Testing, Falsification, and Evaluation

References: [AAAI2026b], [NeurIPS2024], [ACM TOSEM, 2023], [Machine Learning, 2023], [IEEE Trans. Reliability, 2022], [ECCV2022], [Machine Learning, 2021], [ICDM2020], [ICRA2020], [ICSE2019a], [ICSE2019b], [ACM TECS, 2019], [ASE2018], [ACL2023], [AAAI-24]

(d) Robustness Enhancement, Training-Time Guarantees, and Other Properties

References: [IEEE TIFS, 2026], [IEEE T-IFS, 2025], [AAAI-2025c], [CVPR2024], [AAAI2024b], [AAAI2024c], [IEEE Internet of Things, 2024], [Frontier in Neuroscience, 2024], [CVPR2023], [ICCV2023], [IEEE RA-L, 2023a], [AAAI2023], [TMLR, 2022], [CVPR2022], [Frontier in Neuroscience, 2022], [IROS2022b], [ICCV2021], [UAI2021], [ECCV2020a], [NeurIPS2020]

(e) Safety Assurance and Runtime Monitoring

References: [IROS2025], [ICASSP2025], [IROS2024], [ACM TOSEM, 2023], [ACM TECS, 2023], [ITSC2023], [IEEE RA-L, 2023a], [IROS2022b], [DSN2021], [ICCV2021], [SafeCOMP2020]

(f) Foundation Models and AI Agents

References: [ACL2026a], [ACL2026b], [ICLR2026], [AAAI2026a], [AAAI2026b], [CVPR2025], [NeurIPS2025], [ICML2024], [Artificial Intelligence Review, 2024], [AAAI2024a], [ACL2023], [ArXiv, 2023b]

(g) Applications of AI

References: [Pattern Recognition, 2026a], [Pattern Recognition, 2026b], [Expert Systems with Applications, 2026], [Signal Processing, 2026], [AAMAS2023], [ACMMM2023], [IEEE Trans. ITS, 2023], [Pattern Recognition, 2022], [KR2022], [AAMAS2022], [BMVC2021], [IEEE Trans. Medical Imaging, 2021], [ECCV2020b], [MICCAI2020], [IROS2019]

(h) Logic-Based Reasoning and Specification

References: [JLAMP, 2024], [ICFEM2022], [ACM TOCL, 2019], [ACM TOCL, 2018], [IJCAI2018b], [AAAI2017], [IJCAI2017], [AAAI2016a], [AAAI2016b], [IJCAI2016a], [IJCAI2016b], [Artificial Intelligence, 2015], [IJCAI2015], [TACAS2014], [AAAI2014], [KR2014], [AAMAS2013a], [AAMAS2013b], [AAMAS2013c], [AAAI2012a], [AAAI2012b], [IJCAI2011], [AAMAS2010], [ECAI2010]

Open-Source Tools & Software

These aren't just papers — they are open-source tools used by other research groups. All available at github.com/TrustAI.

DLV

DNN verification via SMT. Layer-by-layer exhaustive search for adversarial examples. Foundational tool from our CAV 2017 paper (970+ citations).

DeepConcolic

Concolic testing for DNNs. Structural coverage criteria (neuron, condition, MC/DC), distribution-aware adversarial testing. Extended to RNNs, GNNs, and LLMs.

TrustAI

Open-source toolkit for safety and trustworthiness of deep learning systems. Includes DeepGame, L0-TRE, and other tools from published papers.

MCK

Model checker for verifying autonomous multiagent systems, with support for epistemic and strategic reasoning.

Publications

Google Scholar  |  dblp

TrustAI: Tool Demos

DeepConcolic (Github repository)

Related Publications:

Reliability validation of a learning-enabled dynamic tracking system (Github repository)

Related Publications:

PRODeep: a platform for robustness verification of deep neural networks

testRNN (Github repository)

Related Publications:

Recent Invited Talks, Seminars, and Panel Discussions

Funding & Grants

Robustifying Generative AI through Human-Centric Integration of Neural and Symbolic Methods
Role: PI · EU Horizon · 2025–2028 · More Info · Website
Rare Event Estimation Algorithms for AI Agents
Role: PI · The Alignment Project (UK AI Security Institute / EPSRC) · 2026–2027
ARRES FORECAST: an AI-driven platform to predict road defect evolution
Role: PI (with Robotiz3D) · Innovate UK · 2024–2025 · More Info
Utilising generative AI (LLMs) for searching technical documentation in a cyber-secure environment
Role: PI (with Ronghui Mu) · AKT funded · 2024 · More Info
An Ethical and Robust AI Development Framework: Assessing Correctness and Detecting Fakes
Role: Co-I (PI: Dr Guangliang Cheng) · DSO funded · 2024–2025 · More Info
A literature review on "Safeguarding LLMs"
Role: Co-I (PI: Dr Yi Dong, with Dr Ronghui Mu) · Alan Turing funded · 2024 · More Info
CRoCS: Certified Robust and Scalable Autonomous Operation in Cyber Space
Role: PI (with Alan Marshall, Valerio Selis, Ronghui Mu) · Alan Turing Institute · 2023–2024 · More Info
SPACE: fully decentralised distributed learning for tradeoff of privacy, accuracy, communication complexity, and efficiency
Role: PI (with Xingyu Zhao, Yi Dong) · Innovate UK · 2022–2023 · More Info
SOLITUDE: Safety Argument for Learning-enabled Autonomous Underwater Vehicles
Role: PI (with Xingyu Zhao, Simon Maskell, Sven Schewe, Sen Wang) · Dstl · 2020–2022 · More Info
FOCETA — Foundations for Continuous Engineering of Trustworthy Autonomy
Role: Liverpool lead (with Sven Schewe) · EU H2020 · 2020–2023 · More Info
EnnCore: End-to-End Conceptual Guarding of Neural Architectures
Role: Liverpool lead · EPSRC · 2021–2024 · More Info
MBDA & WSTC project on "Adaptive & reactive mission execution"
Role: PI (with Jason Ralph, Simon Maskell) · 2019–2020
Dstl PhD studentship on "Statistical Approach to Assess the Trustworthiness of Robotics and AI"
Role: PI · 2020–2024
Test Coverage Metrics for Artificial Intelligence — v2.0
Role: PI (with Simon Maskell, Sven Schewe) · Dstl · 2019–2021 · More Info
Test Coverage Metrics for Artificial Intelligence
Role: PI (with Simon Maskell, Sven Schewe, Daniel Kroening) · Dstl · 2018–2019 · More Info
EPSRC ORCA (Offshore Robotics for Certification of Asset) Hub
Co-I (PI at Liverpool: Michael Fisher, with Mike Jump) · EPSRC · 2017–2021
EPSRC 2018 Vacation Bursary Programme project
Role: PI · 2018
KTP Project with Kadfire
Role: Co-I · 2017–2019
Lockey Grant, Oxford University
Role: PI · 2015

Community Leadership

Program Committee Memberships

Key Departmental Administrative Roles

Open Positions

Please check the Home tab for current openings. We have funded positions via RobustifAI (EU Horizon) and AISI (EPSRC).

Alumni Destinations

Our 10+ PhD graduates and postdocs now hold faculty and research positions at Exeter, Southampton, Warwick, Manchester, Imperial College London, MBZUAI, Purple Mountain Laboratories, and other institutions worldwide.

Postdocs / Graduate Research Associates

PhD Students (Primary Supervisor)

If you are interested in doing a PhD in relevant research areas with me, please feel free to contact me. University of Liverpool has a set of established scholarship schemes, including Liverpool CSC award and Sir Joseph Rotblat Alumni Scholarship.

Visitors

The "Robotics and Artificial Intelligence" Reading Group holds a weekly meeting where one member gives a 30–40 minute talk — discussing their own papers, papers from other research groups, or topics of interest. This is followed by Q&A and group discussion.

Membership

Anyone can join by request. If you are interested, please feel free to drop me a message.

Venue & Meeting Time

Meetings are held weekly, both in-person and via Zoom (hybrid). Tuesday 11:00–12:00, UK time.

Talk Schedule

Please refer to the TACPS lab reading group page for detailed schedule information.

Teaching