Jiha Kim 김지하

3D-CNN action recognizers learn appearance and motion in an entangled way, so they lean on static visual cues instead of true motion — which hurts generalization and interpretability on resource-constrained AIoT edge devices. ASTRA is a dual-pathway architecture that separates semantic appearance from pure motion through three components: an STFormer module that models space and time adaptively through cross-attention and estimates frame-wise importance; Importance-Guided Temporal Perturbation, which isolates appearance-free motion by comparing original and perturbed inputs; and a Motion Gated Unit that refines the motion features through dual-gate modulation. Under a leak-free, video-level protocol ASTRA reaches 97.88% on UCF-101 and 81.05% on HMDB-51 with an MViT-v2-S backbone. Against its own R(2+1)D-18 backbone, the motion pathway adds a significant +2.81%p on HMDB-51 (McNemar p < 0.001) while staying equivalent on appearance-saturated UCF-101, at 32.9M parameters, 41.9 GFLOPs and 8.2 ms GPU latency.

Massive machine-type communications congest the random-access channel and burn power on repeated preamble transmissions. DTDO-RA adjusts the backoff indicator (BI) according to the number of transmissions, and uses reinforcement learning — Q-learning and DDPG — to tune the BI together with the maximum number of preamble transmissions (Max Tx) as the network state changes. Widening the backoff range under heavy traffic keeps the procedure succeeding: the success rate, which falls at 10,000 UEs under the standard procedure, holds until 15,000 UEs. At 15,000 UEs a larger BI cuts transmit power by about 98% (3.71 to 0.072 mJ) at the cost of longer access delay, while letting Max Tx widen the backoff range lowers the delay by up to 4.3× relative to BI adjustment. Simulations run to 50,000 UEs and also show that shrinking Max Tx too far raises power rather than saving it.

GANs are widely used to augment scarce data, but a discriminator that leans too hard on the training set overfits, and the generator ends up reproducing images that look like the training images — augmentation that adds nothing. LDGAN opens the discriminator with LIME to show which image regions drive its real/fake decision, and finds that an unconstrained discriminator relies on the whole image, background included. It then limits the discriminator's training with a threshold, set either on the discriminator's mean loss (about 0.5) or on the generator's loss, so that the decision rests on the object rather than the background and the generated images stay diverse. Of 500 images generated per model, 254 from DCGAN show usable eye and nose regions, against 324 and 360 under the two LDGAN limits.

IEEE 802.11bn Ultra High Reliability (UHR) extends multi-link operation toward a reliability-oriented MAC, but multi-link contention under the EMLSR active-link constraint resists closed-form analysis. We develop a Markov/fixed-point closure for three-band MLDs coexisting with single-link legacy stations: exact under saturated, per-link-independent STR; reducing to an availability-based approximation under saturated EMLSR with an explicit validity range; and extending to unsaturated traffic through per-link Bernoulli arrivals. On top of it we build a controller that adapts the per-link minimum contention window from MAC observables inside the standard EDCA/EMLSR loop. Against SLD, STR-MLD and static-EMLSR baselines it raises throughput by 10.6%, lowers 95th-percentile HOL-service delay by 28%, and improves Jain's fairness from 0.910 to 0.954 at 48 stations. Triangulation across the analytical model, a discrete-event simulator and ns-3 closes within a 2–4% mean relative gap.

IEEE 802.11bn defines ultra-high reliability through per-BSS deadline-violation-rate targets. We study these constraints in dense multi-band deployments with heterogeneous AP link sets. LyMAPPO combines independent per-AP PPO with drop-inclusive rate-form Lyapunov duals, one dual-derived congestion price per beacon for cross-BSS coordination, and an auxiliary lower-tail quantile critic. Training and evaluation use an ns-3-in-the-loop environment with 16 APs, three bands, shared-buffer multi-link operation and Markov channels. Across 16 seeds with 110 s episodes, LyMAPPO achieves 15.1/16 feasible BSSs and a network p99 of 0.0017 — 6.1× to 25× lower than three static baselines. The drift-plus-penalty analysis accounts for clipping and establishes a run-time dual-stability condition under the Slater assumption.

STCA-ViT is a hybrid framework integrating a VideoMAE ViT-Base backbone with bidirectional spatio-temporal cross-attention, Sobel-based boundary detection for motion gating, and YOLO-based person attention in parallel. It adds residual contribution fusion with gradient reversal for feature diversity, orthogonal regularization, a multi-head ensemble classifier, and an optional weakly-supervised temporal action localization head that recycles motion features into frame-level scores without temporal annotations. STCA-ViT reaches 90.46% Top-1 on HMDB-51 — above VideoMAE V2 ViT-g (86.08%) — with 98.29% on UCF-101, 79.44% on Kinetics-400 and 65.01% on Something-Something-v2, using 110M parameters and 305 GFLOPs.

Weakly-supervised video anomaly detection localizes frame-level anomalies using only video-level binary labels. Existing methods treat anomaly detection and risk anticipation as independent tasks, discarding cross-task predictive cues. SEAR jointly performs multi-horizon risk anticipation and MIL-based detection through a shared representation via three contributions: a Cross-Task Anticipation Gate, a zero-initialized scalar that modulates detection features with anticipation risk scores; Temporal Contrastive MIL, a hinge-loss objective separating peak anomaly scores from all normal scores; and Curriculum Pseudo-Labeling with Negative Mining. On UCF-Crime, SEAR reaches 85.36% frame-level AUC with frozen dual encoders (~0.59M trainable parameters), and 87.78% AUC under triple-encoder fusion — the highest among frozen-encoder methods. It detects 89% of anomalous videos before annotated event onset.