publications

2026

1. Preprint

   Transferable Multi-Bit Watermarking Across Frozen Diffusion Models via Latent Consistency Bridges

   Hong-Hanh Nguyen-Le, Van-Tuan Tran, Dinh-Thuc Nguyen, and 1 more author

   arXiv preprint arXiv:2506.16600, 2026

   Abs DOI arXiv HTML

   As diffusion models (DMs) enable photorealistic image generation at unprecedented scale, watermarking techniques have become essential for provenance establishment and accountability. Existing methods face challenges: sampling-based approaches operate on frozen models but require costly N-step Denoising Diffusion Implicit Models (DDIM) inversion (typically N=50) for zero-bit-only detection; fine-tuning-based methods achieve fast multi-bit extraction but couple the watermark to a specific model checkpoint, requiring retraining for each architecture. We propose DiffMark, a plug-and-play watermarking method that offers three key advantages over existing approaches: single-pass multi-bit detection, per-image key flexibility, and cross-model transferability. Rather than encoding the watermark into the initial noise vector, DiffMark injects a persistent learned perturbation δat every denoising step of a completely frozen DM. The watermark signal accumulates in the final denoised latent z_0 and is recovered in a single forward pass. The central challenge of backpropagating gradients through a frozen UNet without traversing the full denoising chain is addressed by employing Latent Consistency Models (LCM) as a differentiable training bridge. This reduces the number of gradient steps from 50 DDIM to 4 LCM and enables a single-pass detection at 16.4 ms, a 45\times speedup over sampling-based methods. Moreover, by this design, the encoder learns to map any runtime secret to a unique perturbation at inference time, providing genuine per-image key flexibility and transferability to unseen diffusion-based architectures without per-model fine-tuning. Although achieving these advantages, DiffMark also maintains competitive watermark robustness against distortion, regeneration, and adversarial attacks.

2. AAAI

   Beyond Binary Classification: A Semi-supervised Approach to Generalized AI-generated Image Detection

   Hong-Hanh Nguyen-Le, Van-Tuan Tran, Dinh-Thuc Nguyen, and 1 more author

   In AAAI Conference on Artificial Intelligence, 2026

   DOI arXiv HTML

2025

1. ACM CSUR

   A survey on proactive deepfake defense: Disruption and watermarking

   Hong-Hanh Nguyen-Le, Van-Tuan Tran, Thuc Nguyen, and 1 more author

   ACM Computing Surveys, 2025

   DOI HTML
2. IJCAI

   Think Twice before Adaptation: Improving Adaptability of DeepFake Detection via Online Test-Time Adaptation

   Hong-Hanh Nguyen-Le, Van-Tuan Tran, Thuc Nguyen-Dinh, and 1 more author

   In International Joint Conference on Artificial Intelligence (IJCAI), 2025

   DOI arXiv HTML
3. Preprint

   FLAME: Towards Federated Fine-Tuning Large Language Models Through Adaptive SMoE

   Khiem Le, Van-Tuan Tran, Ting Hua, and 1 more author

   arXiv preprint arXiv:2506.16600, 2025

   DOI arXiv HTML
4. MCDC@ICLR

   Revisiting Sparse Mixture of Experts for Resource-adaptive Federated Fine-tuning Foundation Models

   Van-Tuan Tran, Khiem Le, and Quoc-Viet Pham

   In ICLR Workshop on Modularity for Collaborative, Decentralized, and Continual Deep Learning (MCDC@ICLR), 2025

   Abs HTML

   Existing federated fine-tuning methods for large-scale foundation models (FMs) assign heterogeneous low-rank adaptation (LoRA) ranks for clients based on their computation capabilities to address system heterogeneity. However, these approaches require merging LoRA matrices into the original model to obtain the full model, causing the computational overhead for resource-constrained clients at inference time. Moreover, their performance is not as effective as that of the homogeneous LoRA, in which the lowest rank is applied to all clients. To overcome these limitations, we propose a resource-adaptive federated fine-tuning method by revisiting the conditional computation property of Sparsely-activated Mixture-of-Experts (SMoE). The key principle here is to extend the data-conditional computation property of SMoE to a new dimension - resource-conditional computation, where clients can activate a suitable number of experts depending on their available resources. Furthermore, to address the imbalanced expert utilization caused by heterogeneous expert activation patterns, we propose a new Activation-aware aggregation algorithm for SMoE (A3SMoE). This algorithm enhances the aggregation process by incorporating client-specific expert activation patterns. Through experiments across independent and identically distributed (IID) and non-IID scenarios, we demonstrate that our proposed method achieves superior performance compared to both homogeneous- and heterogeneous-LoRA approaches under different computation budgets. We also show that LoRA-based methods can be improved when integrated with A3SMoE.

5. ECAI

   ToFU: Transforming How Federated Learning Systems Forget User Data

   Van-Tuan Tran, Hong-Hanh Nguyen-Le, and Quoc-Viet Pham

   In European Conference on Artificial Intelligence (ECAI), 2025

   DOI arXiv HTML

2024

1. Preprint

   Passive deepfake detection across multi-modalities: A comprehensive survey

   Hong-Hanh Nguyen-Le, Van-Tuan Tran, Dinh-Thuc Nguyen, and 1 more author

   arXiv preprint, 2024

   DOI arXiv HTML
2. IJCNN

   D-captcha++: A study of resilience of deepfake captcha under transferable imperceptible adversarial attack

   Hong-Hanh Nguyen-Le, Van-Tuan Tran, Dinh-Thuc Nguyen, and 1 more author

   In International Joint Conference on Neural Networks (IJCNN), 2024

   Abs DOI HTML

   The advancements in generative AI have enabled the improvement of audio synthesis models, including text-to-speech and voice conversion. This raises concerns about its potential misuse in social manipulation and political interference, as synthetic speech has become indistinguishable from natural human speech. Several speech-generation programs are utilized for malicious purposes, especially impersonating individuals through phone calls. Therefore, detecting fake audio is crucial to maintain social security and safeguard the integrity of information. Recent research has proposed a D-CAPTCHA system based on the challenge-response protocol to differentiate fake phone calls from real ones. In this work, we study the resilience of this system and introduce a more robust version, D-CAPTCHA++, to defend against fake calls. Specifically, we first expose the vulnerability of the D-CAPTCHA system under the transferable imperceptible adversarial attack. Secondly, we mitigate such vulnerability by improving the robustness of the system by using adversarial training in D-CAPTCHA deepfake detectors and task classifiers.

3. IEEE TETC

   Personalized privacy-preserving framework for cross-silo federated learning

   Van-Tuan Tran, Huy-Hieu Pham, and Kok-Seng Wong

   IEEE Transactions on Emerging Topics in Computing, 2024

   Abs DOI HTML

   Federated learning (FL) is recently surging as a promising decentralized deep learning (DL) framework that enables DL-based approaches trained collaboratively across clients without sharing private data. However, in the context of the central party being active and dishonest, the data of individual clients might be perfectly reconstructed, leading to the high possibility of sensitive information being leaked. Moreover, FL also suffers from the nonindependent and identically distributed (non-IID) data among clients, resulting in the degradation in the inference performance on local clients’ data. In this paper, we propose a novel framework, namely Personalized Privacy-Preserving Federated Learning (PPPFL), with a concentration on cross-silo FL to overcome these challenges. Specifically, we introduce a stabilized variant of the Model-Agnostic Meta-Learning (MAML) algorithm to collaboratively train a global initialization from clients’ synthetic data generated by Differential Private Generative Adversarial Networks (DP-GANs). After reaching convergence, the global initialization will be locally adapted by the clients to their private data. Through extensive experiments, we empirically show that our proposed framework outperforms multiple FL baselines on different datasets, including MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100.

4. Preprint

   Deepfake generation and proactive deepfake defense: A comprehensive survey

   Hong-Hanh Nguyen-Le, Van-Tuan Tran, Dinh-Thuc Nguyen, and 1 more author

   2024

   Abs HTML

   The proliferation of highly realistic deepfakes, powered by Generative Artificial Intelligence (GenAI), presents significant challenges to digital trust and security. This survey provides a comprehensive overview of proactive deepfake detection approaches, including disruption and watermarking methods. Our survey provides a taxonomy of these strategies based on their existing methodologies and extend the discussion to other perspectives, including imperceptibility, transferability, universality, and robustness. We also explore the associated threat models, considering various adversary objectives and capabilities. Additionally, we review state-of-the-art deepfake generation techniques that provide context for the challenges faced by detection methods.

2023

1. IEEE Access

   Learning from multiple expert annotators for enhancing anomaly detection in medical image analysis

   Khiem H Le, Van-Tuan Tran, Hieu H Pham, and 3 more authors

   IEEE Access, 2023

   Abs DOI HTML

   Recent years have experienced phenomenal growth in computer-aided diagnosis systems based on machine learning algorithms for anomaly detection tasks in the medical image domain. However, the performance of these algorithms greatly depends on the quality of labels since the subjectivity of a single annotator might decline the certainty of medical image datasets. In order to alleviate this problem, aggregating labels from multiple radiologists with different levels of expertise has been established. In particular, under the reliance on their own biases and proficiency levels, different qualified experts provide their estimations of the "true" bounding boxes representing the anomaly observations. Learning from these nonstandard labels exerts negative effects on the performance of machine learning networks. In this paper, we propose a simple yet effective approach for the enhancement of neural networks’ efficiency in abnormal detection tasks by estimating the actually hidden labels from multiple ones. A re-weighted loss function is also used to improve the detection capacity of the networks. We conduct an extensive experimental evaluation of our proposed approach on both simulated and real-world medical imaging datasets, MED-MNIST and VinDr-CXR. The experimental results show that our approach is able to capture the reliability of different annotators and outperform relevant baselines that do not consider the disagreements among annotators.