UQ Cyber and AUSCERT to host the International Cybersecurity Challenge
The 2026 International Cybersecurity Challenge will be held on the Gold Coast from 18–21 May, bringing global cyber talent to Australia.
Event page →
University of Queensland · School of EECS
Detection you can defend. Explanations you can act on.
Research across deepfake detection, multimodal misinformation, audio forensics, and AI security.
About
Applied forensic AI for high-stakes digital evidence
Led by Dr Priyanka Singh, Senior Lecturer in Cyber Security at The University of Queensland, the group works across cyber security, digital forensics, privacy and security, homomorphic encryption, cloud computing and AI-enabled content authentication.
Our core problem is urgent: synthetic and manipulated media is becoming easier to generate, harder to detect and more persuasive when paired with misleading captions or narratives. The lab studies the full chain from manipulation and propagation to detection, explanation, provenance and trustworthy user-facing verification.
We focus on systems that can survive real deployment: robust models, auditable evidence, privacy-aware workflows and outputs that people outside a benchmark setting can understand.
News
Recent news and public updates
Recent public commentary, cyber security engagement and lab updates.
ABC News Hour commentary on the Canvas data breach
Dr Singh discussed the global Canvas breach affecting education institutions and the importance of calm, evidence-based public cyber security communication.
Expert commentary on AI-generated fake digital IDs
ABC News quoted Dr Singh on the rise of AI-generated fake digital IDs and the risks created when synthetic identity documents become easy to produce.
Read ABC story →How to spot a deepfake video
UQ Contact asked Dr Singh to explain deepfake risks and practical verification cues for public audiences in the Sora 2 era.
Read guide →What even are internet cookies?
UQ Contact featured Dr Singh in a plain-language explainer on cookies, tracking, login persistence and online privacy.
Read story →Research
Research areas
Six active research directions across multimedia forensics, network security, audio deepfake detection and AI safety. Select a card to view figures and context from the underlying papers.
Multimodal manipulation & misinformation detection
Manipulation is increasingly a communication problem, not just a pixel-level artefact problem. Persuasive posts can pair a real image with a misleading caption, a generated image with a true caption, or a coordinated package of image, text and external context. Our work treats detection as a region-aware and evidence-aware reasoning task.
- SGS (Segmentation-Guided Scoring) extends HAMMER to handle global scene inconsistencies, such as a politician giving a speech in front of an erupting volcano, by separating foreground and background regions and comparing their semantic descriptions.
- D-SECURE combines a content-based manipulation detector (HAMMER) with a retrieval-based fact-checker (DEFAME) so the same pipeline can catch both subtle local edits and globally fabricated claims.
Explainable deepfake detection
DF-P2E reframes deepfake detection as an explanation pipeline. Instead of stopping at a confidence score, it combines a classifier, Grad-CAM-style visual evidence, image-to-text captioning and LLM-based narrative refinement so users can inspect the reasons behind a verdict and ask follow-up questions about it.
- The classifier produces the initial real/fake decision and confidence.
- A captioning module describes manipulated regions in forensic-aware language.
- An LLM refinement layer adapts the explanation to non-expert audiences (journalists, investigators, legal users).
Cross-domain audio deepfake detection
Audio deepfake detectors trained on one dataset routinely fail on another because of differences in recording conditions, codecs and synthesis methods. We study lightweight, interpretable pipelines that adapt detectors across domains without target-domain labels.
- Pre-trained Wav2Vec 2.0 embeddings provide a strong self-supervised front end.
- A modular sequence (power transform, ANOVA-based feature selection, joint PCA, and CORAL alignment) narrows the gap between source and target distributions.
- The pipeline is transparent and ablatable, making each step's contribution easy to audit before deployment.
Explainable network intrusion detection
Network intrusion detection systems often deliver near-perfect benchmark numbers but offer little for the analyst on the other end of an alert. eX-NIDS adds an LLM-based explanation layer that turns flagged NetFlows into human-readable justifications, augmented with cyber threat intelligence and protocol context.
- The Prompt Augmenter extracts IP threat intelligence, geo-localisation, IP history and protocol IDs for each malicious flow.
- This context is fused with a base prompt, then sent to an LLM (Llama 3 / GPT-4) to generate explanations.
- A novel evaluation framework checks the explanations for correctness, factual consistency and feature consistency.
LLM safety & adversarial misuse
Two complementary directions: controlling LLM behaviour at the activation level for trustworthy generation, and studying how non-experts can already turn off-the-shelf chatbots into phishing tools. Both inform how cyber teams should reason about LLMs as both a defence surface and an attack surface.
- PIXEL performs adaptive, position-wise activation steering with a property-aligned subspace, providing a tuning-free way to align attributes such as truthfulness.
- Chatbot phishing study shows how participants without prior security or coding knowledge can prompt-engineer ChatGPT, Copilot and Gemini to produce phishing emails that consistently outperform manually-written ones.
Benchmarks & dataset foundations
Detectors are only as honest as the benchmarks they're trained and evaluated on. We build dataset extensions and evaluation suites that expose blind spots in current state-of-the-art models, particularly global scene plausibility and end-to-end factuality, which mainstream multimodal manipulation datasets do not cover.
- DGM4+ extends the DGM⁴ benchmark with 5,000 high-quality samples introducing foreground–background mismatches and hybrids with text manipulation (FG-BG, FG-BG+TA, FG-BG+TS).
- ClaimReview2024+ alongside DGM⁴ in D-SECURE supplies factuality labels (supported / refuted / NEI) so detection and fact-checking can be evaluated jointly.
- Failure-case curation drives method design: an illustrative example (e.g., a politician in front of a volcano) motivates the entire region-aware reasoning direction.
Privacy-preserving forensicsEncrypted-domain analysis, homomorphic image hashing and secure matching for sensitive institutional media.
Cyber threat intelligence sharingSecure collaborative CTI workflows that reduce privacy leakage while improving incident response.
Knowledge distillation for LLMsCompressing and adapting massive language models for embedded, industrial and domain-specific deployment.
Publications
Publications and technical themes
Selected themes across the group's work
The group's public research profiles span digital forensics, multimedia forensics, encrypted-domain processing, cloud security, privacy-preserving AI and deepfake detection.
Digital forensics
Multimedia forensics
Encrypted-domain processing
Deepfake detection
Privacy-preserving AI
AI safety
Network intrusion detection
Knowledge distillation
UQ Experts profile →
Workshops
Workshops
Selected workshops organised or co-organised across multimedia forensics, digital forensics, AI security and digital public infrastructure.
PP-MisDet: From Perception to Persuasion
From Perception to Persuasion: Challenges and Advances in Misinformation Detection in Society. A CVPR 2026 workshop on multimodal misinformation, falsified visual evidence, trustworthy evaluation and high-stakes societal impact.
Workshop page →IMPACT: Impact-aware Multimodal Persuasive Analysis and Contextual Trust
An ACM Multimedia 2026 workshop on multimedia-native, persuasion-optimised misinformation, contextual trust, provenance, robustness and human-centred verification.
Workshop page →Understanding and Analysing Executable and Linkable Format (ELF) Files
A DFRWS APAC 2024 workshop by Parag Rughani and Priyanka Singh on Linux binary internals, ELF file structure and malicious ELF analysis for digital forensics.
Workshop page →Chain of Infection Detection: Cross-Domain Forensic Artefact Correlation
A hands-on DFRWS USA 2026 workshop on reconstructing infection chains by correlating artefacts across volatile memory, persistent storage and network captures.
Workshop page →Workshop on AI Integration in Businesses: A Perspective on Efficiency VS Security Risks
A UQ–University of Exeter collaboration examining how AI adoption improves automation, decision-making and productivity while creating new cyber security, privacy and model-risk challenges.
Workshop page →Digital Public Infrastructure and Public Service Delivery: Inclusion, Efficiency, and Security
A UQ–Indian Council for Research on International Economic Relations collaboration under the Australia–India Cyber and Critical Technology Partnership, funded by DFAT, focused on DPI-enabled public service delivery.
Workshop page →Transformation of Health Systems through Digital Public Infrastructure
A UQ–Indian Council for Research on International Economic Relations collaboration under the Australia–India Cyber and Critical Technology Partnership, funded by DFAT, focused on digital public infrastructure for health systems.
Workshop outcome document →Gallery
Workshops, talks and community moments
Selected photos from lab activities, public engagement and cyber forensics community events.
DFRWS APAC 2024 · organisers and keynote speakers
DFRWS APAC 2024 · workshop and volunteers
AI integration in businesses workshop
QuASI panel · AI governance
Queensland Government AI Executive Leadership Program
ANU National Youth Science Forum
SDLT 2025 · AI-powered distributed ledgers
UQ Young Alumni · deepfakes mini masterclass
Team
People and mentoring
Leadership and student team across multimodal misinformation, audio deepfake detection and cyber forensics.
Dr Priyanka Singh
Senior Lecturer in Cyber Security · School of EECS · Affiliate, UQ Cyber Research Centre
Multimedia forensics
Deepfake detection
Privacy-preserving AI
Lab lead. Research spans cyber security, digital forensics, privacy-preserving AI, homomorphic encryption and AI-enabled content authentication, with public commentary on deepfakes, synthetic identity and online verification.
UQ Experts profile →
Gagandeep Singh
Research student · School of EECS · T-IMPACT initiative
Multimodal misinformation
Scene plausibility
Benchmarks
I'm a research student in the School of Electrical Engineering and Computer Science at the University of Queensland. My work focuses on multimodal misinformation detection and extending state-of-the-art detectors. I also build benchmarks and pipelines that combine content-based manipulation detection with retrieval-based fact-checking.
LinkedIn →
Samudi Amarasinghe
Master's student · School of EECS
Misinformation
Cybersecurity
Machine learning
I am a Master's student studying computer science and cybersecurity in the School of Electrical Engineering and Computer Science at the University of Queensland. My area of interest is misinformation, with a focus on evidence combination for detection. I also strive to improve the explainability of detection, to make systems accessible to all users.
LinkedIn →
Urawee Thani
Software Engineering thesis student · School of EECS
Machine learning
Audio deepfake detection
Cross-domain generalisation
I'm a software engineering student currently undertaking a thesis in machine learning, speech processing, and audio deepfake detection. My recent work focuses on cross-domain generalisation and improving the robustness of AI models across diverse datasets and real-world conditions.
LinkedIn →Contact
Contact and collaboration
Affiliation
The University of Queensland
School of Electrical Engineering and Computer Science
Brisbane, Australia