22 publications · 12 accepted · 10 preprints & under review. Also on Google Scholar.
Journal of Visual Communication and Image Representation, vol. 120, art. 104895, 2026
First-author journal article
Summary Proposed GSCNet, a transformer-based style-transfer network with granular control over style injection that suppresses the artifacts of global style-transfer pipelines.
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV ’27), Algorithms Track, 2027
Summary Proves a self-correction theorem for quantized diffusion sampling (Wasserstein-2 error bounded by the sensitivity-weighted sum of per-step quantization errors) and derives a water-filling mixed-precision bit allocation that explains the early-coarse / late-fine rule; on a pretrained CIFAR-10 DDPM it is the most faithful allocation to the full-precision sampler at every budget.
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV ’27), Evaluations and Datasets Track, 2027
Summary Diagnoses why zero-cost NAS proxies can rank against accuracy when transported across search spaces: on NAS-Bench-201 a dominant capacity-associated factor explains 52.9%/49.4% of a 29-feature proxy zoo’s variation, and several raw proxies reverse sign (condition-number sum +0.88 → −0.77); two controlled search traps cost +5.40 and +1.97 points at K=10.
Proceedings of the International Conference on Digital Image Computing: Techniques and Applications (DICTA ’26), 2026
Accepted; first-author work
Summary Proposed a scene-level conformal calibration rule for co-visibility scores with finite-sample connectivity guarantees; obtained 15/17 connected proxy-positive test graphs, versus 11/17 for a fixed threshold and 2/17 for top-30% selection.
Proceedings of the British Machine Vision Conference (BMVC ’26), 2026
First-author work — complementary defense to the WWW ’26 attack paper
Summary Proposed SLICE, which injects watermark signals into compartmentalized regions of the semantic latent space; led model design, theoretical analysis and evaluation, with statistical false-accept-rate guarantees and robustness to semantic-guided regeneration attacks.
Proceedings of the ACM International Conference on Multimedia (ACM MM ’26), 2026
Summary Contributed core algorithmic design and experimental validation of a layered, cross-modal watermarking framework that embeds tamper-sensitive signals across audio and visual streams, enabling deepfake detection and segment-level localization.
Proceedings of the European Conference on Computer Vision (ECCV ’26), 2026
Summary Contributed to a bi-level prompting framework that transfers and completes cross-modal information, sustaining visual-recognition accuracy when input modalities are partially missing.
Proceedings of the ACM Web Conference 2026 (WWW ’26, Short Paper Track), Dubai, United Arab Emirates, April 13–17, 2026, pp. 8569–8572
Summary Identified that the semantic-binding mechanism of content-aware semantic watermarking can be bypassed by globally coherent semantic rewrites, and proposed CSI, an LLM-guided attack that invalidates state-of-the-art schemes while keeping image fidelity. Led the work end-to-end.
2024 3rd International Conference on Artificial Intelligence and Computer Information Technology (AICIT), Yichang, China, IEEE, 2024
Sole-authored undergraduate work
Summary Proposed G-RXAD, a deep-reinforcement-learning-based model for network / system attack detection.
NeurIPS 2026 Workshop on AI for Stochastic Dynamics (STODY), 2026
* Equal first-author contribution
Summary Co-led a training-free attack that stochastically deflects the reverse denoising trajectory to remove diffusion watermarks; evaluated against nine watermarking methods with 95–100% reported attack success.
NeurIPS 2026 Workshop on Beyond Private Training: The New Landscape of AI Privacy (InfPriv), 2026
Summary Contributed to an exact worst-case privacy analysis of retry-until-valid wrappers around pure ε-differentially-private inference routines, with a closed-form worst-case envelope that stays bounded away from zero as the attempt cap grows.
NeurIPS 2026 Workshop on Continual Learning in the Era of Foundation Models and Embodied Agents, 2026
Summary Contributed to a study of how much testing is needed to certify that an agent update retains earlier capabilities; a frozen model-selection update raised GSM8K accuracy by 11.0 points with the tested historical policy preserved.
arXiv preprint arXiv:2610.08137, 2026
First-author work — conceptual research agenda
Summary First-authored a research agenda comparing detection and watermarking for visuals from generative models versus LLM-written code, and posed ten scoped research questions.
arXiv preprint arXiv:2609.12418, 2026
* Equal contribution
Summary Contributed to RiPPLE, which ranks whole NAS search spaces by extrapolating early-training learning curves of a few anchor architectures; evaluated on twelve benchmark cells plus DARTS.
arXiv preprint arXiv:2609.11884, 2026
* Equal contribution
Summary Contributed to CoRA-NAS, a coarse-to-fine NAS method at about 1% of the full-training budget that reaches Spearman 0.715–0.946 across four NAS benchmarks without fully trained labels.
arXiv preprint arXiv:2609.10416, 2026
Summary Contributed to TrajMark, a training-free symmetric-key watermark for coding-agent READ actions; it recovered the exact owner in all clean batches and detected 95.5–100% of eligible single-site edits.
arXiv preprint arXiv:2609.08213, 2026
* Equal first-author contribution
Summary Co-developed DRIFT, a black-box attack that removes diffusion watermarks by partial forward diffusion followed by stochastic reverse resampling; 98–100% attack success across nine watermarks in three paradigms.
arXiv preprint arXiv:2608.03539, 2026
arXiv preprint arXiv:2607.13003, 2026
Summary Contributed to an information-theoretic account of what can be recovered about an embedded watermark from observed outputs.
arXiv preprint arXiv:2607.08400, 2026
First-author work — public project page with interactive demo
Summary Proposed TRACE, which stamps two complementary watermark channels onto an LLM agent’s decision trajectory so attribution survives deleted or rewritten steps. Led design, implementation and evaluation.
arXiv preprint arXiv:2606.14688, 2026
Summary Contributed to a theoretical study showing, in the language-generation-in-the-limit framework, that large volumes of low-value output are provably necessary for generating valuable mathematics.
arXiv preprint arXiv:2510.01910, 2025
External collaboration with KAIST
Summary Co-designed RoGRAD, the first iterative paradigm that uses RAG to inject class-consistent augmentations with iterative graph contrastive learning, yielding up to 82.4% average improvement over conventional GNN- and LLM-enhanced baselines.
No papers match. Clear the search or pick another topic.