Technical Summary: Resource-Efficient WiFi CSI Sensing via Exploiting the Age of Samples

Problem Statement

WiFi Channel State Information (CSI) sensing is a powerful technique for inferring human activity and identity without wearable devices. However, existing high-accuracy models typically assume dense, regular, and high-rate CSI sampling. In practical scenarios, particularly with the standardization of Integrated Sensing and Communication (ISAC) in IEEE 802.11bf, sensing must coexist with data communication. This coexistence imposes a strict average sensing budget, limiting the fraction of time slots available for acquiring fresh CSI measurements. Furthermore, packet loss and traffic-driven irregularities result in non-uniform sampling intervals.

Current approaches to sparse or irregular CSI generally fall into three categories:

1. Reconstruction methods: Attempt to recover dense CSI from sparse data before inference, introducing computational complexity and potential reconstruction errors.
2. Resource-selection methods: Prune subcarriers or antennas but often ignore temporal sampling rates.
3. Time-aware attention models (e.g., UniFi): Embed absolute timestamps and interpolate data onto a uniform grid. However, these methods do not explicitly evaluate accuracy as a function of an average sensing budget, nor do they utilize the Age of Information (AoI)—the time elapsed since a sample was captured—as a trainable feature.

Consequently, there is a need for a sensing framework that explicitly accounts for sample freshness (AoI) under strict resource constraints without incurring significant computational overhead.

Methodology

System Model and Sampling Constraints

The authors formulate the problem within a measurement session of $T$ discrete time slots. An average sensing budget $\alpha \in (0, 1]$ constrains the fraction of slots where CSI samples are acquired. Two fixed sampling policies are considered to model different practical regimes:

• Deterministic (Accumulated/Bresenham) Policy: Spreads $M$ sensed slots near-uniformly over the session, modeling scheduled dedicated sensing.
• Stochastic (Bernoulli) Policy: Samples slots independently with probability $\alpha$, modeling random packet loss and traffic-driven irregularities.

The "age" of the $m$-th sensed frame, $\delta(t_m)$, is defined as the time elapsed since its capture relative to the decision instant ($T-1$): $\delta(t_m) = (T-1) - t_m$. This metric is treated as a zero-overhead side signal available at the MAC layer.

Proposed Framework: Age-Aware WiFi Sensing

The authors propose a low-cost, age-aware framework that explicitly incorporates sample freshness into model training. The architecture consists of:

1. Packed Observation: Instead of carrying stale frames through unsensed slots, the input is a packed set of $M$ sensed frames, each tagged with its age: $X = \{(X(t_m), \delta(t_m))\}_{m=1}^M$.
2. CSI Encoder: A convolutional neural network (CNN) that processes the packed CSI tensor (antennas $\times$ subcarriers $\times$ frames) to produce a motion-content embedding ($e_{csi}$). Adaptive average pooling ensures the model accepts variable numbers of frames ($M$) while outputting a fixed-size embedding.
3. Age Encoder: A lightweight branch that maps scalar age values $\delta(t_m)$ to a reliability embedding ($e_a \in (0, 1)^d$). This is achieved via a shared linear layer, batch normalization, and a sigmoid activation, followed by averaging over the session. The output represents a per-dimension reliability weight based on staleness.
4. Multiplicative Fusion: The CSI and age embeddings are fused via an element-wise (Hadamard) product: $e_f = e_{csi} \odot e_a$. This design treats age as a "freshness-dependent gate," scaling CSI features based on their reliability rather than treating age as an independent modality.
5. Classifier: A post-fusion block (ReLU, LayerNorm, Dropout) maps the fused embedding to class scores.

The framework adds negligible parameters (only $4d$ additional scalars for $d=128$) and maintains computational complexity comparable to a standard CSI-only model.

Key Contributions

• Problem Formulation: The paper formalizes WiFi CSI sensing under an average sensing budget constraint, explicitly modeling both scheduled sparse sensing and random packet loss.
• Age-Aware Architecture: It introduces a novel, low-overhead framework that encodes the Age of Information (AoI) of CSI samples and fuses it multiplicatively with CSI features. This allows the model to dynamically attenuate features from stale samples.
• Zero-Overhead Implementation: The approach utilizes timing information already available at the MAC layer, requiring no extra signaling or complex reconstruction steps.
• Comprehensive Evaluation: The method is evaluated on two distinct tasks (Human Activity Recognition and Person Identification) using the NTU-Fi datasets under both deterministic and stochastic sampling regimes.

Experimental Results

The proposed model was evaluated on the NTU-Fi Human Activity Recognition (HAR) and NTU-Fi HumanID datasets. It was benchmarked against:

1. A CSI-only baseline.
2. The state-of-the-art UniFi (mTAN) time-aware attention model.
3. A Last-M baseline (using only the most recent $M$ frames).

Key Findings:

• Superior Performance: The age-aware model consistently outperformed both the CSI-only baseline and the UniFi benchmark across all sampling regimes and budget levels.
• Low-Budget Gains: The performance improvement was most significant under strict sensing budgets. For Person Identification, the proposed model achieved up to a 10-percentage-point improvement over the UniFi method. Under accumulated sampling at $\alpha=0.01$, the age-aware model retained 80.7% accuracy, whereas the CSI-only baseline dropped to 72.7% and UniFi to 68.6%.
• Temporal Coverage vs. Recency: Experiments comparing "Accumulated" sampling (spread over time) vs. "Last-M" (concentrated at the end) revealed that covering the entire session is more critical than merely using recent frames. At equal frame counts, the accumulated approach significantly outperformed the Last-M approach.
• Task Sensitivity: The benefits of age-awareness were more pronounced in Person Identification (a temporally sensitive task requiring gait pattern resolution) than in Activity Recognition.

Significance and Claims

The paper claims that freshness-aware inference is an effective and practical approach for maintaining sensing accuracy when CSI availability is limited or irregular. By explicitly modeling the "age" of samples, the proposed framework adapts to resource constraints without the latency or complexity of data reconstruction.

The authors emphasize that their approach is particularly valuable in the context of IEEE 802.11bf, where sensing and communication must share the medium. The ability to maintain high accuracy even at very low sensing budgets (e.g., $\alpha=0.01$) demonstrates that the proposed age-aware mechanism effectively mitigates the degradation caused by sparse and irregular sampling, offering a resource-efficient solution for real-world ISAC deployments.