Organizations: Basque Center for Applied Mathematics (BCAM), 48009 Bilbao, Spain · College of Semiconductor (College of Integrated Circuits), Hunan University, Changsha 410082, China · College of Computer Science, Beijing Information Science and Technology University, Beijing 100192, China · Department of Electronic Engineering and the Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing 100084, China · Shanghai AI Laboratory, Shanghai 201112, China
Environmental identification in wireless sensing is essential for 6G integrated sensing and communication (ISAC) systems to achieve reliable situational awareness. However, deep learning (DL) models for this task often fail to generalize under domain shift across diverse environments. While the Inter-Instance Variational Auto-encoder (IIns-VAE) learns features of rich representation, its neural classifier remains vulnerable to these distribution changes. In this paper, we propose IIns-VAE+, a hybrid model that combines the IIns-VAE framework with Minimax Risk Classifiers (MRC) to improve adaptability in transfer learning scenarios. We use real-world datasets to evaluate our framework across three transfer learning scenarios, including general to specific room environments, high to low label resolutions, and mixed to specific environments. The experimental results indicate that IIns-VAE+ significantly outperforms baselines, demonstrating its critical value in building adaptable and robust perceptive networks in future 6G systems.
Figures & tables
Fig. 1: Transfer learning in environmental identification. Target domain data can be collected from different environments compared to source domain data.
Category
Core Idea
Representative Model
References
1. Based on Convolutional Neural Network (CNN)
Treats CSI as an image (2D grid data) and uses CNN to extract its spatial features (e.g., beam direction, multipath structure).
CSILoc : Uses 1D CNN to process multi-channel (amplitude, phase) CSI, followed by global average pooling.
[ 37 ]
Duloc : Employs a dual-channel CNN, processing CSI divided into two branches based on subcarrier stability.
[ 38 ]
2. Based on Recurrent Neural Network (RNN)
Uses RNNs (e.g., LSTM, GRU) to process temporal features in CSI, capturing how the channel varies over time.
Zhang et al. : Uses an LSTM network to model preprocessed CSI amplitude and phase information, extracting temporal features.
[ 39 ]
3. Based on Attention Mechanism
Uses self-attention mechanisms to fuse information from different antennas, subcarriers, and time dimensions in CSI, better capturing long-range dependencies.
LoT : Leverages the Vision Transformer (ViT) concept, converting CSI into an image format processed by a standard Transformer model.
[ 40 ]
Swin-Loc : Uses Swin Transformer layers to extract features from processed CSI data.
[ 41 ]
ACPNet : Combines two types of attention mechanisms (self-attention and channel attention) to process CSI from MIMO systems.
[ 42 ]
TABLE I: Summary of Deep Learning-Based Indoor Localization Algorithms
Category
Core Idea
Representative Work(s)
Reference(s)
1. Foundational Theories
1.1 Variational Inference
Approximate complex posterior distributions by optimizing a tractable variational distribution.
Variational Inference Framework
[ 44 ]
1.2 Deep Generative Models
Learn complex data distributions using deep neural networks for generation and representation learning.
Variational Autoencoder (VAE)
[ 45 ]
Generative Adversarial Network (GAN)
[ 46 ]
1.3 Implicit Distributions
Model complex distributions without explicit density functions through stochastic transformations.
Implicit Distribution Models
[ 47 ]
2. Localization under Uncertainty
TABLE II: Summary of Variational Inference for Wireless Sensing and Localization
Fig. 2: Comparison of the neural classifier and the MRC classifier under domain shift. The MRC classifier provides resilience to domain shift and ensures more adaptability to new data domain.
Fig. 3: Graphical model for IIns-VAE+ with the environment latent variables for the source and target domains.
Fig. 4: Network structures of layer-wise and bottleneck-wise integration strategies.
Fig. 5: Comparison of PCA Feature Clustering Across Databases.
Model Type
Method
Accuracy
Precision
Recall
F1 Score
Baselines (No Transfer)
PFs + SVM
0.840
0.853
0.840
0.845
PFs + MRC
0.834
0.875
0.834
0.843
IIns-VAE
0.541
0.351
0.351
0.351
Layer Integration
IIns-VAE
0.818
0.809
0.627
0.651
IIns-VAE+
0.875
0.883
0.875
0.864
Bottleneck Integration
IIns-VAE
0.829
0.790
0.676
0.705
TABLE III: Performance Comparison for General to Specific Room (LOS/NLOS Classification Task).
Model Type
Method
Accuracy
Precision
Recall
F1 Score
Baselines (No Transfer)
PFs + SVM
0.240
0.499
0.240
0.295
PFs + MRC
0.267
0.394
0.267
0.249
IIns-VAE
0.135
0.027
0.200
0.048
Layer Integration
IIns-VAE
0.420
0.199
0.276
0.223
IIns-VAE+
0.524
0.465
0.524
0.486
Bottleneck Integration
IIns-VAE
0.387
0.189
0.219
0.143
TABLE IV: Performance Comparison for High to Low Label Resolutions ( 5 Material Labels Classification Task).
Method Type
Method
Accuracy
Precision
Recall
F1 Score
Baselines (No Transfer)
PFs + SVM
0.489
0.394
0.489
0.410
PFs + MRC
0.457
0.351
0.457
0.378
IIns-VAE
0.383
0.176
0.210
0.132
Layer Integration
IIns-VAE
0.323
0.124
0.181
0.124
IIns-VAE+
0.510
0.536
0.500
0.503
Bottleneck Integration
IIns-VAE
0.526
0.330
0.384
0.336
TABLE V: Performance Comparison for Mixed to Specific Environment ( 5 Material Labels Classification Task).
Fig. 6: Performance comparison for the three models in terms of the accuracy, precision, recall, and F1 score. The results illustrate the better adaptability of IIns-VAE+ models, especially IIns-VAE+ (Layer), in handling diverse environments with limited target domain data.
Fig. 7: Comparison of Accuracy for IIns-VAE with MRCs under layer and bottleneck integration setting across different transfer learning scenarios.
Fig. 8: Comparison of F1 Score for IIns-VAE with MRCs under layer and bottleneck integration setting across different transfer learning scenarios.
Augmented Cognition Meta-communications ERC Research Center, Korea University, Korea · Department of Computer Science, Munster Technological University (MTU), Ireland · School of Computing, Faculty of Engineering and Computing, Dublin City University, Ireland +4
School of Electrical Engineering and Computer Science, National University of Sciences and Technology, Islamabad, Pakistan · Department of Electronic Engineering, Kyung Hee University, Yongin 17104, Republic of Korea