ODAI@UOS - Research

[Arxiv] CR-QAT: Curriculum Relational Quantization-Aware Training for Open-Vocabulary Object Detection

Extreme low-bit (e.g., 4-bit) quantization of open-vocabulary object detection (OVOD) models severely degrades fine-grained vision-language alignment and distorts inter-region relational structures. To resolve this, we propose Curriculum Relational Quantization-Aware Training (CR-QAT), which mitigates error accumulation through progressive model partitioning and preserves multi-dimensional structural knowledge via text-centric relational distillation. Experiments on LVIS and COCO zero-shot benchmarks demonstrate that CR-QAT consistently outperforms existing baselines under aggressive low-bit settings, achieving relative AP improvements of up to 40.9%.

[DAC'26] Adaptive Spiking Neural Networks for Real-Time Multi-Object Detection Tasks

The core principle of RT-SNN is to provide resource-efficient execution options for SNNs with varying execution times, ensuring timing guarantees while considering their impact on accuracy. RT-SNN employs a scheduling framework that dynamically selects the optimal execution option for accuracy while maintaining timing guarantees. Deployed on Spiking-YOLO, an SNN-based MOD model from ANN-to-SNN conversion, RT-SNN demonstrates effectiveness in meeting timing and accuracy requirements while improving energy efficiency.

[Arxiv] STAS: Spatio-Temporal Adaptive Computation Time for Spiking Transformers

STAS is designed to dynamically adjust the number of tokens processed during inference in SNN-based vision transformers (ViTs), considering power consumption proportional to the number of tokens. It extends the applicability of Adaptive Computation Time (ACT), previously limited to RNNs and standard ViTs, to SNN-based ViTs by selectively discarding less informative spatial tokens, thereby improving efficiency.

[Arxiv] SPOILER: TEE-Shielded DNN Partitioning of On-Device Secure Inference with Poison Learning

Existing TEE-shielded DNN partitioning methods fail to balance privacy and efficiency, suffering from either intrinsic leakage or high latency caused by structural dependencies. To address this tradeoff, we propose SPOILER, a novel search-before-training framework that leverages hardware-aware neural architecture search (NAS) to structurally decouple the TEE sub-network, maximizing parallel execution. Furthermore, SPOILER establishes rigorous logical isolation via self-poisoning learning, rendering the exposed backbone functionally incoherent without the secured TEE component.

[Arxiv] Zero2Text: Zero-Training Cross-Domain Inversion Attacks on Textual Embeddings

Vector databases in RAG introduce critical privacy risks via embedding inversion. Existing methods fail in black-box settings due to prohibitive query costs or reliance on in-domain data. To address this, we propose Zero2Text, a training-free framework using recursive online alignment with dynamic ridge regression. We demonstrate that standard defenses, including differential privacy, remain ineffective against this adaptive threat.

[AAAI'26] Timestep-Compressed Attack on Spiking Neural Networks through Timestep-Level Backpropagation

The timestep-compressed attack (TCA) is a novel framework that addresses the high latency of adversarial attacks on Spiking Neural Networks (SNNs) by leveraging their unique operational properties. It introduces timestep-level backpropagation for early stopping and reuses the initial membrane potential to eliminate redundant computations, significantly accelerating attack generation.

[IJCAI'25] BankTweak: Adversarial Attack against Multi-Object Trackers by Manipulating Feature Banks

BankTweak targets the feature extractor during the association phase, exposing vulnerabilities in the Hungarian matching method commonly used in feature-based MOT systems. By strategically injecting altered features into the feature banks without modifying object positions, BankTweak induces persistent ID switches, enhancing attack efficiency and robustness even after the attack ends.

[RTSS'25] CF-DETR: Coarse-to-Fine Transformer for Real-Time Object Detection

CF-DETR leverages a coarse-to-fine detection Transformer architecture and a real-time scheduling framework to dynamically adjust inference based on object criticality. This approach guarantees timely detection of safety-critical objects while selectively improving overall accuracy, effectively managing the latency-accuracy trade-off in autonomous perception systems.

[EMSOFT'24] Batch-MOT: Batch-Enabled Real-Time Scheduling for Multi-Object Tracking Tasks

Batch-MOT is the first system designed to achieve both (G1) timing guarantees and (G2) accuracy maximization by leveraging batch execution, enabling multiple DNN executions within a single inference. It accomplishes G1 and G2 without significant runtime overhead by systematically exploiting runtime execution behaviors enabled by the adaptive framework.

[RTSS'22] RT-MOT: Confidence-Aware Real-Time Scheduling Framework for Multi-Object Tracking Tasks

RT-MOT is a novel system design for multi-object tracking (MOT) that addresses the dual requirements of timely execution and high tracking accuracy under limited computing resources. Utilizing a confidence-aware scheduling framework and predictive tracking accuracy estimation optimizes workload pair selection for detection and association, ensuring offline timing guarantees while maximizing real-time performance.

[NRF 미래도전연구] Real-Time SNN-LMM for Neuromorphic AI Semiconductors (25.9~30.8)

This project develops a low-power Spiking Neural Network (SNN) based on a Large Multi-modal Model (LMM), supported by a custom neuromorphic AI semiconductor and a system software platform for real-time applications.
Through a novel co-design of the model, hardware, and software, we aim to simultaneously achieve high accuracy, low power consumption, and real-time performance.

[ITRC] Low Earth Orbit (LEO) Satellite Systems (23.7~30.12)

We aim to develop a lightweight satellite reconnaissance system that simultaneously achieves real-time performance and high accuracy for multi-purpose LEO satellites.