TY - GEN
T1 - Robust HAR with CSI Tokens
T2 - 2026 International Conference on Computing, Networking and Communications, ICNC 2026
AU - Miao, Fucheng
AU - Takyu, Osamu
AU - Shan, Lin
AU - Zhao, Ou
AU - Ohtsuki, Tomoaki
AU - Gui, Guan
AU - Adachi, Fumiyuki
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - This paper presents a WiFi Channel State Information (CSI)-based human activity recognition framework that fuses class-aware textual prompts with CSI patch' tokens through a Prompt-Query CrossModal Attention front end. A lightweight 'patch reprogramming' layer aligns CSI tokens to the languageembedding space, after which a frozen GPT-2 performs contextual modeling; only small task-specific heads are trained. To improve representation learning without altering inference, three optional training-only auxiliaries, supervised contrastive learning, masked reconstruction, and temporal order prediction-act as complementary regularizers. The overall objective is optimized as a simple weighted sum without uncertainty weighting. Experiments on two public CSI datasets show consistent gains over representative Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Residual Network (ResNet), and Vision Transformer (ViT) baselines, faster and smoother convergence, robustness under Signal-to-Noise Ratio (SNR) degradation (where SNR refers to the quality of the received WiFi signal in typical indoor multipath environments), and cleaner confusion matrices on hard class pairs. These results suggest that injecting textual priors and aligning CSI tokens before using a frozen LLM is an effective way to create stronger, noise-resilient CSI representations.
AB - This paper presents a WiFi Channel State Information (CSI)-based human activity recognition framework that fuses class-aware textual prompts with CSI patch' tokens through a Prompt-Query CrossModal Attention front end. A lightweight 'patch reprogramming' layer aligns CSI tokens to the languageembedding space, after which a frozen GPT-2 performs contextual modeling; only small task-specific heads are trained. To improve representation learning without altering inference, three optional training-only auxiliaries, supervised contrastive learning, masked reconstruction, and temporal order prediction-act as complementary regularizers. The overall objective is optimized as a simple weighted sum without uncertainty weighting. Experiments on two public CSI datasets show consistent gains over representative Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Residual Network (ResNet), and Vision Transformer (ViT) baselines, faster and smoother convergence, robustness under Signal-to-Noise Ratio (SNR) degradation (where SNR refers to the quality of the received WiFi signal in typical indoor multipath environments), and cleaner confusion matrices on hard class pairs. These results suggest that injecting textual priors and aligning CSI tokens before using a frozen LLM is an effective way to create stronger, noise-resilient CSI representations.
KW - Deep Learning
KW - WiFi sensing
KW - channel state information (CSI)
KW - human activity recognition (HAR)
KW - large language models (LLMs)
KW - multi-task regularization
KW - noise robustness
UR - https://www.scopus.com/pages/publications/105035733174
UR - https://www.scopus.com/pages/publications/105035733174#tab=citedBy
U2 - 10.1109/ICNC68183.2026.11416929
DO - 10.1109/ICNC68183.2026.11416929
M3 - Conference contribution
AN - SCOPUS:105035733174
T3 - 2026 International Conference on Computing, Networking and Communications, ICNC 2026
BT - 2026 International Conference on Computing, Networking and Communications, ICNC 2026
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 16 February 2026 through 19 February 2026
ER -