Skip to content
Cyber Science LabCanada Cyber Foundry, University of Guelph, Canada
2027

The First International Workshop on AI Forensics (AIF 2027)

May 20, 2027 Le Centre Sheraton Montreal, Canada

Co-located with the 48th IEEE Symposium on Security and Privacy (IEEE S&P 2027). Call for papers; submission deadline 12 February 2027 (AoE).

About the workshop

What does it take for evidence about a machine-produced artifact to hold up, whether that artifact is an AI system or its output?

The incidents are already here: agents that take unintended actions, models that leak training data, automated decisions contested in litigation, synthetic media offered as evidence of something that never occurred. The methods and evidentiary standards needed to investigate any of this to a defensible standard do not yet exist.

Security research has produced a large literature on attacks against machine learning systems and their defences. Far less work addresses investigation after a deployed AI system has been compromised. Unlike mature areas of digital forensics, AI forensics lacks widely accepted investigative methods, validated tools, evidentiary standards and case law. Regulators now require incident reporting, litigants require explanation, and courts are already receiving synthetic media as exhibits.

AIF brings together researchers and practitioners working on the forensic investigation of AI systems and the forensic examination of AI-generated content. The two communities are converging on the same evidentiary problem from opposite directions, and the workshop puts them in one room.

Workshop tracks

Track 1: Forensic investigation of AI systems

Model runtimes, inference servers, agents and fine-tuning pipelines all leave state. What survives, what can be recovered, and what can be inferred from it?

  • Memory forensics of model runtimes and inference servers; recovery of prompts, contexts and residual state
  • Reconstruction of agent behaviour from execution traces, tool-call records and orchestration logs
  • Forensic artifacts of agentic frameworks and protocols
  • Model identification and provenance; establishing which model and configuration produced an output
  • Forensic soundness of acquisition from executing systems; atomicity and integrity of snapshots
  • Forensic readiness: logging, retention and attestation designed to support later investigation
  • Investigation of fine-tuning, data poisoning and model supply-chain incidents
  • Anti-forensics against AI-system investigation

Track 2: Forensic examination of AI-generated content

Detection and attribution results stated as evidentiary claims: what is asserted about a specific artifact, under what conditions, with what error rates, and where the method fails.

  • Attribution of synthetic images, video, audio and text to generating models or families
  • Detection under adversarial conditions and distribution shift
  • Watermarking, fingerprinting and content provenance as forensic evidence
  • Error rates, validation and reliability characterization
  • Examination of partially synthetic and edited media
  • Datasets and benchmarks for evaluating content forensics

Track 3: Evidence

Admissibility, chain of custody and expert practice apply identically whether the artifact under examination is a model or a video.

  • Admissibility of AI-derived evidence; expert-witness practice
  • Chain of custody and integrity for machine-produced artifacts
  • Standards for AI forensic investigation, including extensions of existing forensic ontologies and process standards
  • Regulatory requirements bearing on investigability
  • Case studies from real investigations

What we are looking for

AIF is a research workshop at a top-tier security venue. We want work that makes and defends a claim about what can be established from an AI system or its output: what a method demonstrates, under what threat model, with what error characteristics, and where it fails. Papers should state what is being established and with what confidence, and evaluate the limits of the claim. In Track 2, conditions should include the realistic recompression, resizing, transcoding and laundering any real artifact undergoes before it reaches an examiner. Work that establishes the limits of an existing method is as welcome as work that improves its accuracy. Enumerations of artifacts with no accompanying claim about soundness or attribution are better suited to other venues.

Submission guidelines

All accepted papers appear in the IEEE Security and Privacy Workshops proceedings, published by IEEE Computer Society Press and available through IEEE Xplore.

Research papers

13 pages of text. Plus up to 5 pages for references and appendices, totaling no more than 18 pages. Previously unpublished work with a stated claim and its evaluation.

SoK papers

13 pages of text. Up to 5 pages for appendices; references do not count towards the number of pages. Title prefix “SoK:” and a checkbox on the submission form.

Short papers

6 pages of text. Plus up to 5 pages for references and appendices. Preliminary results, position arguments, or practitioner case studies. A workshop category with no main-conference equivalent.

Systematization of Knowledge. We solicit SoK papers that evaluate, systematize, and contextualize existing knowledge in AI forensics. Suitable papers provide an important new viewpoint on an established research area, support or challenge long-held beliefs with compelling evidence, or present a convincing, comprehensive new taxonomy. Survey papers without such insights are not appropriate and may be rejected without full review. SoK papers are held to the same standards as research papers, but are accepted based on their treatment of existing work and value to the community, not on new research results.

Page limit and formatting

  • All text and figures past page 13 (page 6 for short papers) must be clearly marked as part of the appendix. Reviewers are not required to read appendices.
  • Papers that are over the allowed number of pages will be rejected without review. The final camera-ready paper must be no more than 18 pages.
  • US letter (not A4) paper, using the IEEE “compsoc” conference proceedings template. LaTeX submissions must use IEEEtran.cls version 1.8b: \documentclass[conference,compsoc]{IEEEtran}. See the “IEEE Demo Template for Computer Society Conferences” on Overleaf. There is no matching MS Word template.
  • Papers that use the non-compsoc IEEE conference template, modify margins, font or line spacing, or use egregious space scrunching are subject to rejection without review. Some packages (e.g. \usepackage{usenix}) override compsoc formatting and must be removed.
  • Submissions must be in PDF.

Anonymous submission

  • No author names or affiliations, real or the default fake names in the IEEE template, may appear on the title page. Avoid revealing identity in the text and acknowledgments (e.g. funding, colleagues who gave feedback).
  • Cite your own previous work in the third person. Where that is infeasible, blind the reference itself.
  • Anonymize artifact repositories (e.g. GitFront or Anonymous GitHub), including account names, repository names and comments. Repositories must not be updated after the deadline.
  • Do not include full CVE identifiers.
  • You may give talks, post a preprint to arXiv, and disclose vulnerabilities to vendors, but refrain from widely advertising results. Do not contact PC members about your submission.
  • Papers that are not properly anonymized may be rejected without review.

Author policies

AIF follows the policies of the IEEE S&P 2027 call for papers, adapted only where the workshop format differs.

Originality and simultaneous submission. All submissions must be original work. Authors must clearly document any overlap with previously published or simultaneously submitted papers from any of the authors; failure to point out and explain overlap is grounds for rejection. Simultaneous submission of the same paper to another venue with proceedings or a journal is not allowed and will be grounds for automatic rejection. Papers rejected from the IEEE S&P 2027 main conference are welcome.

Conflicts of interest. Authors must declare all and only their conflicts with program committee members in HotCRP at submission. A conflict exists when the PC member: is a co-author of the paper; has been a co-worker in the same company or organization within the past two years; has been affiliated with the same academic institution as a co-author within the past two years (PhD students are conflicted with the university they graduated from for two years after graduation); has collaborated on a research paper, including ongoing or unpublished work, within the past two years; has been a collaborator on a funding grant within the past two years, whatever its status; is or was the author’s primary thesis advisor or advisee, no matter how long ago; or is a relative or close personal friend. Explain any other conflict in the corresponding HotCRP field. Papers with incorrect or incomplete conflict information are subject to rejection, and changes to the author list are not permitted after submission. Program co-chairs with a conflict are excluded from handling that paper.

Research ethics considerations. All papers must complete the “Ethics Considerations” field in HotCRP. Authors of accepted papers will add an Ethics Considerations section to the camera-ready, where it does not count toward page limits. If there are none, write “None” with a short explanation of how you reached that conclusion. Papers that report no ethics considerations where the program committee believes they are necessary may be rejected.

Work identifying vulnerabilities should disclose them to affected vendors, and the submission should describe the steps taken or planned. Work involving human subjects or data derived from them should state whether it received institutional ethics review (e.g. IRB) approval or a waiver and how participants were treated, and should detail harm mitigation for any personally identifiable or sensitive data. Work interacting with live systems not owned by the authors should explain permission and harm mitigation. Work developing new tools, technologies or attacks should discuss potential misuse, mitigation, and how benefits were weighed against harms. Authors are encouraged to consult the Menlo Report. Ethical review is especially relevant to AI forensics work that handles recovered prompts, user conversations, or media depicting real people.

Generative AI policy. AIF follows the IEEE policy on generative AI and the IEEE S&P 2027 policy. Authors may use generative AI as an object of study and when preparing their paper, but must disclose and motivate that use in the “Generative AI usage considerations” field in HotCRP. Authors are responsible for the accuracy and originality of all content, including the literature review. Where generative AI was used to improve writing, state: “Generative AI was used for editorial purposes in this manuscript, and all outputs were inspected by the authors to ensure accuracy and originality.” Where it is integral to the methodology, detail its use and the limitations it introduces, such as reproducibility with closed models, and justify the computational footprint of experiments. An AI cannot be listed as an author. Non-compliance is grounds for desk rejection.

Competing interests, authorship and ORCID. Authors must disclose relevant financial and non-financial competing interests in the camera-ready version, beyond standard funding lines (for example, an advisory role with or a position at a company working in the area of the paper). Changes to the author list (adding, removing or reordering) are not permitted after submission. All authors must provide an ORCID identifier at submission, using an email and full name that match those in HotCRP.

Withdrawal, presentation and deadlines. A paper can be withdrawn at any point before reviews have been sent to the authors, but not afterwards. Authors are responsible for obtaining appropriate publication clearances, and at least one author of each accepted paper must register and present the paper in person at the workshop. Authors should have a draft submitted 24 hours before the deadline; submissions that fail because the server crashed within 24 hours of the deadline will not be accepted. All deadlines are 23:59:59 AoE (UTC−12).

Important dates

  • Paper submission: 12 February 2027 (Anywhere on Earth)
  • Author notification: 19 March 2027 (By email from HotCRP)
  • Camera-ready: 8 April 2027 (Aligned with the main conference)
  • Workshop: 20 May 2027 (Le Centre Sheraton, Montreal)

All deadlines are 23:59:59 Anywhere on Earth (UTC−12). The submission deadline falls three and a half weeks after the IEEE S&P 2027 early-reject notifications of 18 January 2027.

Submission deadline: 12 February 2027 (AoE). IEEE “compsoc” template, US letter, PDF. Have a draft in 24 hours before the deadline.

Tentative schedule

  • 09:00 · Opening remarks
  • 09:10 · Keynote 1
  • 09:55 · Coffee break
  • 10:20 · Session 1: Track 1 papers
  • 12:00 · Lunch
  • 13:00 · Keynote 2
  • 13:45 · Session 2: Track 2 papers
  • 15:05 · Coffee break
  • 15:30 · Closing panel
  • 16:20 · Closing remarks

Times are Eastern Time and indicative.

Organizing committee

Steering committee

  • Chair: Ibrahim Baggili, Louisiana State University, USA
  • Mauro Conti, University of Padua, Italy and Örebro University, Sweden
  • Anderson Rocha, University of Campinas, Brazil
  • Brendan Saltaformaggio, Georgia Institute of Technology, USA
  • Matthew Stamm, Drexel University, USA
  • Mourad Debbabi, Concordia University, Canada

Program committee

  • Aisha Ali-Gombe, Louisiana State University, USA
  • Haider Al-Khateeb, Aston University, UK
  • Giovanni Apruzzese, Reykjavik University, Iceland
  • Wauter Bosma, Netherlands Forensic Institute, The Netherlands
  • Rainer Böhme, Universität Innsbruck, Austria
  • Alessandro Brighente, University of Padua, Italy
  • Eoghan Casey, University of Lausanne, Switzerland
  • Marco Centenaro, University of Padua, Italy
  • Kam-Pui Chow, University of Hong Kong, China
  • William Corrias, Sapienza University of Rome, Italy
  • Fabio De Gaspari, Sapienza University of Rome, Italy
  • Steven H. H. Ding, McGill University, Canada
  • Denis Donadel, Fondazione Bruno Kessler, Italy
  • Isao Echizen, National Institute of Informatics, Japan
  • Felix Freiling, Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany
  • Benjamin C. M. Fung, McGill University, Canada
  • Teddy Furon, Centre Inria de l’Université de Rennes, France
  • Alberto Giaretta, Örebro University, Sweden
  • Yong Guan, Iowa State University, USA
  • Mohammad Hammoudeh, King Fahd University of Petroleum and Minerals, Saudi Arabia
  • Sokratis Katsikas, Norwegian University of Science and Technology, Norway
  • Ryan K. L. Ko, University of Queensland, Australia
  • Nhien-An Le-Khac, University College Dublin, Ireland
  • David Oygenblik, Georgia Institute of Technology, USA
  • Jungheum Park, Korea University, Republic of Korea
  • Nikolaos Pitropakis, Deree, The American University of Greece, Greece
  • Alessandro Piva, University of Florence, Italy
  • Indrakshi Ray, Colorado State University, USA
  • Golden G. Richard III, Louisiana State University, USA
  • Mark Scanlon, University College Dublin, Ireland
  • Simone Soderi, University of Florence, Italy
  • Issa Traore, University of Victoria, Canada
  • Stefano Tubaro, Politecnico di Milano, Italy
  • Lingyu Wang, University of British Columbia Okanagan, Canada
  • Junichi Yamagishi, National Institute of Informatics, Japan
  • Li Yang, Ontario Tech University, Canada

Contact

For questions about scope or submissions, write to the program co-chairs, Paria Shirani and Sajad Homayoun. For general enquiries, contact the general chair, Ali Dehghantanha.

The workshop takes place at Le Centre Sheraton Montreal Hotel, the venue of the main conference (17–20 May 2027). Registration and hotel information are handled by the main conference. See the IEEE S&P 2027 website.