2026·CKAIA 2026
Simple Pseudobulk Outperforms Topological and Quantum-Geometric Readouts of Single-Nucleus Chromatin and Transcriptomes for Donor-Level Alzheimer's Classification: A Rigorous Benchmark in a Powered Multiome Cohort
Sunjun Hwang, Dohyun Hwang
Topological data analysis and quantum-geometric distances have been proposed as donor-level readouts of single-cell disease state, on the premise that the geometry and shape of a cell-state manifold within a cell type carry information that simple pseudobulk averaging discards. We tested this premise rigorously in a powered, openly available Alzheimer's disease multiome cohort with 111 donors, 6 brain regions, and over 2.2 million transcriptomic and 1.2 million chromatin nuclei. We first established that donor-level pathology is genuinely detectable: RNA pseudobulk classified late-AD versus non-AD donors at AUC 0.86, and chromatin pseudobulk reached AUC 0.91 in excitatory neurons, with consistent positive controls. We then benchmarked persistent-homology features under classical and quantum-geometric distances, on both a standard latent and a learned geometry-preserving autoencoder latent. Under repeated cross-validation and covariate adjustment, simple pseudobulk-mean was the most robust readout across all six cell types; topological and quantum-geometric features performed at or near chance and never robustly exceeded pseudobulk. A single apparent win did not replicate under repeated cross-validation and collapsed to chance after sequencing-depth adjustment, illustrating how single-split evaluation and depth confounding can manufacture false topological signal. We conclude that parsimonious pseudobulk is the appropriate baseline for this task, and that topological readouts require this baseline, repeated cross-validation, and depth control before any claim of added value.
Single-Cell GenomicsAlzheimer's DiseaseTopological Data AnalysisPseudobulkBenchmark
2026·IEEE COINS 2026Accepted
H2C: A Pan-Cancer Gene Panel Discovered via Persistent Homology in Topological Autoencoder Latent Space
Sunjun Hwang, Dohyun Hwang, Eunho Choi
RNA-seq-based cancer biomarker discovery predominantly relies on Euclidean statistics such as t-tests and fold-change analysis, which cannot capture multivariate nonlinear interactions among genes. We propose a pipeline combining Topological Autoencoders (TAE) with Persistent Homology (PH) to analyze the topological structure of gene expression data. Applied to TCGA-BRCA (1,215 samples, 20,862 genes), our pipeline encodes RNA-seq profiles into a 32-dimensional latent space with topology-preserving constraints and computes Vietoris–Rips persistent homology. Size-matched permutation tests show that tumor samples exhibit approximately 2.5× more H1 loops than normal samples (p < 0.001). Through decoder Jacobian-based gene traceback, we identify 37 genes (the H2C panel) that rank within the TDA top 200 yet are not significant under standard univariate analysis. H2C achieves strong classification performance (AUC = 0.993 for BRCA, AUC = 0.977 for pan-cancer across 33 cancer types) and significant prognostic value (log-rank p = 2.03×10⁻⁷). These results suggest TDA provides a complementary perspective for identifying cancer-associated signals beyond standard Euclidean analysis.
Topological Data AnalysisPersistent HomologyCancer GenomicsAutoencoderRNA-seq
2026·ICUFN 2026Accepted
AdaFed: Adaptive Selective Aggregation for Heterogeneous Federated Learning in Autonomous Driving
Sunjun Hwang, Dohyun Hwang
We present a comprehensive empirical study of federated learning (FL) for seven heterogeneous autonomous driving models and propose AdaFed, an adaptive selective aggregation strategy. We introduce a tier-based partial backbone sharing framework that groups models by shared architectural components (ResNet-34/50 backbones, BEV encoders, and BEV-Lift modules) and evaluate six FL strategies: FedAvg, FedProx, SCAFFOLD, FedMD, FedDF, and AdaFed. Across all seven models, no FL strategy surpasses the per-model imitation-learning baseline; the practical question is therefore which strategy minimises degradation. AdaFed applies a tier-specific base strategy, blends aggregated and local weights with an adaptive coefficient, and reverts any round-level update that worsens validation error. Experiments on CARLA 0.9.16 reveal that partial backbone averaging stays closest to the IL baseline for ResNet-50 models, knowledge-distillation approaches degrade performance, and AdaFed yields the best FL result for the two models most exposed to negative transfer — VAD and UniAD. These findings provide practical guidelines for heterogeneous FL deployment in autonomous driving.
Federated LearningAutonomous DrivingHeterogeneous ModelsAdaptive AggregationCARLA
2026·ICAIIC 2026·Tokyo, Japan
Adversarial Robustness Analysis of Deep Learning-Based Automatic Modulation Classification in Wireless Communication
Sunjun Hwang, Eunho Choi, Dohyun Hwang
Deep learning-based Automatic Modulation Classification (AMC) has emerged as a critical technology for spectrum sensing and cognitive radio applications. However, the vulnerability of these models to adversarial attacks poses significant security concerns in wireless communication systems. This paper presents a comprehensive evaluation of adversarial robustness for a VTCNN2-based AMC model using the RadioML2016.10A dataset. We systematically analyze three representative adversarial attacks — FGSM, DeepFool, and C&W — and evaluate two defense mechanisms: adversarial training and Denoising Autoencoder (DAE). Our experimental results demonstrate that the baseline model is highly susceptible to adversarial perturbations, with accuracy dropping from 54.02% to as low as 10.80% under DeepFool attack. FGSM-based adversarial training improves robustness across multiple attacks, with a modest clean-accuracy drop. DAE preprocessing preserves clean accuracy but provides only limited gains under iterative L2 attacks. These findings highlight the urgent need for robust defense mechanisms in deep learning-based wireless communication systems.
Adversarial RobustnessSecurityWireless CommunicationDeep Learning
