WiMoT: Replay-Free Continual Wi-Fi Sensing via Moment Transport

Chunchun Yang; Yang Cao; Zhitang Chen

cs.AIOct 23, 2026

WiMoT: Replay-Free Continual Wi-Fi Sensing via Moment Transport

Authors: Chunchun Yang, Yang Cao, Zhitang Chen

Organizations: School of Information Science and Technology, University of Science and Technology of China, Hefei 230026, China · Independent Researcher

Abstract

Continual Wi-Fi sensing must balance plasticity, learning new rooms or users, against stability, keeping earlier ones, yet nodes often cannot retain channel state information (CSI) for privacy and storage reasons. We propose WiMoT (Wi-Fi Moment Transport), which stores the per-class feature mean and covariance of every earlier domain instead of samples. When the representation changes, stored statistics become stale. WiMoT restricts the change to an affine map on a frozen feature basis and applies the same map to the statistics, so they stay valid without old samples. Supervised contrastive learning learns the new domain, a closed-form loss on the stored statistics keeps earlier classes separable, and a Gaussian discriminant classifier, refitted from the transported statistics with equal weight per domain, predicts without domain identity. On 3DO, WiMANS, and Widar3.0, WiMoT's average final accuracies are 94.02%, 89.61%, and 95.16% with at most 0.58 percentage points (pp) of forgetting. On 3DO and Widar3.0, it also has the highest balance of new- and earlier-domain accuracy and is 3.77 and 13.11 pp ahead of the strongest of 21 baselines, which uses replay, staying ahead across domain orders and label budgets. On WiMANS, it matches the strongest replay baseline in accuracy (+0.41 pp) while reducing forgetting from 2.78 to 0.00 pp, and has the highest six-order average accuracy. Because full covariances dominate the memory, storing them as rank-32 factors shrinks the moment memory to 6.4%, losing at most 0.49 pp of accuracy, within edge-node storage budgets.

Figures & tables

Panel of Figure 1 — Fig. 1: WiMoT framework. (a) A frozen encoder, the semantic stem, the first-domain PCA shortcut, and a fixed random amplitude stem form the basis $\phi$ . (b) The affine adapter maps $\phi$ to $h$ , from which the OSCL classifier predicts $\hat{y}$ (bold path); the projector with SupCon and the linear map $W$ with $\mathcal{L}_{\mathrm{hist}}$ are used only in training. (c) Class moments of each domain are stored before its samples are deleted; they supply $B_{s},K_{s}$ to $\mathcal{L}_{\mathrm{hist}}$ and, transported through the current adapter, the moments used by OSCL.

Methods without replay (no historical individual is stored)
Method	3DO		WiMANS		Widar3.0
	AFA (%) $\uparrow$	AFR (pp) $\downarrow$	AFA (%) $\uparrow$	AFR (pp) $\downarrow$	AFA (%) $\uparrow$	AFR (pp) $\downarrow$
WiMoT	$\boldsymbol{94.02\pm 0.07}$	$0.32\pm 0.33$	$\boldsymbol{89.61\pm 0.47}$	$\boldsymbol{0.00\pm 0.46}$	$\boldsymbol{95.16\pm 1.23}$	$\boldsymbol{0.58\pm 0.58}$
Fine-tuning	$84.19\pm 1.13$	$17.62\pm 1.70$	$86.93\pm 1.17$	$4.48\pm 0.96$	$70.02\pm 0.38$	$28.78\pm 0.34$
EWC (2017) [ 17 ]	$84.19\pm 1.10$	$17.62\pm 1.66$	$86.93\pm 1.17$	$4.48\pm 0.96$	$70.62\pm 0.37$	$28.19\pm 0.38$
SI (2017) [ 18 ]	$89.38\pm 0.21$	$9.42\pm 0.29$	$87.76\pm 1.52$	$4.01\pm 0.96$	$75.51\pm 0.28$	$20.17\pm 0.36$

TABLE I: Continual Wi-Fi sensing performance (mean

\pm

sample SD, three runs). Bold: best of each column; WiMoT is also bold when it ranks second

Stream	Domains (order)	Labels	Train/val/test	Channels	Cache
3DO	3 days	3 activities	2400/600/1200	52	240
WiMANS	3 rooms	9 activities	378/108/108	270	36
Widar3.0	9 users, Room 1	6 gestures	446–450/149–150/150	540	42

TABLE II: Data streams (counts per domain; Cache: capacity of the replay cache)

Panel of Figure 4 — Fig. 3: Mean test accuracy over the observed domains after each training stage (original order, three-run means)

Method	Balance	AFR	New	Earlier	Balance	AFR	New	Earlier	Balance	AFR	New	Earlier
	3DO				WiMANS				Widar3.0
WiMoT	$\boldsymbol{94.94\pm 0.20}$	0.32	93.07	96.82	$\boldsymbol{89.12\pm 0.61}$	0.00	87.50	90.74	$\boldsymbol{95.69\pm 0.77}$	0.58	95.08	96.31
Fine-tuning	$87.62\pm 0.82$	17.62	96.49	78.76	$87.42\pm 0.94$	4.48	88.89	85.96	$81.56\pm 0.17$	28.78	95.28	67.83
SI+ER	$91.92\pm 0.54$	5.64	93.60	90.24	$88.27\pm 0.35$	4.01	89.66	86.88	$87.18\pm 0.59$	13.67	92.94	81.42
MAS+ER	$90.97\pm 0.75$	5.21	91.71	90.24	$\boldsymbol{89.27\pm 0.67}$	2.78	90.12	88.43	$80.01\pm 0.80$	10.47	83.31	76.72
LwF+ER	$88.69\pm 0.30$	$\mathbf{-1.17}$	83.94	93.44	$88.12\pm 0.74$	1.23	88.43	87.81	$86.31\pm 0.40$	10.25	89.78	82.83

TABLE III: Balance, New, Earlier (%), and AFR (pp) of WiMoT and selected baselines (mean

\pm

sample SD, seeds 42–44). Bold: best of each column and dataset among all twenty-two methods; WiMoT is also bold when it ranks second

	3DO			WiMANS			Widar3.0
Configuration	AFA	$\Delta$ AFA	AFR	AFA	$\Delta$ AFA	AFR	AFA	$\Delta$ AFA	AFR
WiMoT (full)	$\boldsymbol{93.98\pm 0.15}$	–	$\boldsymbol{0.26}$	$\boldsymbol{89.57\pm 0.55}$	–	$-0.28$	$\boldsymbol{95.24\pm 0.88}$	–	$0.42$
WiMoT w/o historical moments	$76.94\pm 2.44$	$-17.04\pm 2.58$ (0/5)	$30.54$	$84.26\pm 1.05$	$-5.31\pm 1.44$ (0/5)	$5.00$	$48.15\pm 1.11$	$-47.10\pm 1.07$ (0/5)	$52.45$
WiMoT w/o SupCon	$91.44\pm 0.77$	$-2.54\pm 0.77$ (0/5)	$\boldsymbol{-0.57}$	$84.44\pm 2.61$	$-5.12\pm 2.31$ (0/5)	$\boldsymbol{-3.89}$	$92.31\pm 0.66$	$-2.93\pm 1.07$ (0/5)	$\boldsymbol{0.15}$
WiMoT w/o $\mathcal{L}_{\mathrm{hist}}$	$93.89\pm 0.43$	$-0.09\pm 0.44$ (2/5)	$0.62$	$88.52\pm 0.86$	$-1.05\pm 0.56$ (0/5)	$0.83$	$94.73\pm 0.86$	$-0.52\pm 0.73$ (1/5)	$1.10$
WiMoT, frozen adapter ( $t>1$ )	$92.92\pm 0.43$	$-1.07\pm 0.32$ (0/5)	$1.33$	$86.23\pm 0.60$	$-3.33\pm 0.86$ (0/5)	$0.09$	$89.32\pm 0.51$	$-5.93\pm 0.59$ (0/5)	$0.40$

TABLE IV: Single-component ablation (five runs): AFA (%), paired AFA difference from full WiMoT (pp; number of seeds with higher AFA), and AFR (pp). Bold: best AFA and lowest AFR of each dataset (the control row is not ranked); the full model is also bold when it ranks second

	3DO		WiMANS		Widar3.0
Configuration	AFA	Earlier	AFA	Earlier	AFA	Earlier
R0 Frozen adapter ( $t>1$ )	$92.98\pm 0.47$	$95.39\pm 0.46$	$86.17\pm 0.59$	$87.96\pm 0.73$	$89.35\pm 0.51$	$90.88\pm 0.56$
R1 SupCon, not transported	$81.91\pm 1.66$	$74.36\pm 2.65$	$85.12\pm 0.55$	$84.72\pm 0.73$	$63.78\pm 2.51$	$61.03\pm 2.55$
R2 SupCon $+\,\mathcal{L}_{\mathrm{hist}}$ , not transported	$78.62\pm 1.53$	$68.94\pm 2.09$	$85.93\pm 1.49$	$85.83\pm 1.52$	$66.73\pm 2.37$	$64.45\pm 2.49$
R3 SupCon, transported	$93.96\pm 0.46$	$96.62\pm 0.57$	$88.52\pm 0.86$	$89.35\pm 0.65$	$94.79\pm 0.88$	$95.70\pm 0.70$
R4 SupCon $+\,\mathcal{L}_{\mathrm{hist}}$ , transported (full WiMoT)	$\boldsymbol{94.01\pm 0.14}$	$\boldsymbol{96.78\pm 0.30}$	$\boldsymbol{89.57\pm 0.55}$	$\boldsymbol{90.83\pm 0.21}$	$\boldsymbol{95.23\pm 0.90}$	$\boldsymbol{96.40\pm 0.81}$

TABLE V: Transport ladder: AFA and Earlier accuracy (%, mean

\pm

sample SD, seeds 42–46) with and without transport of the stored moments. Bold: best of each column

AFA (%) $\uparrow$
	Room 1, nine users				Room 2	Room 3
Method	Original	Reverse	4-8-9-2-3-5-1-7-6	6-7-4-3-1-2-5-8-9	2 users	4 users
WiMoT	$\boldsymbol{95.16\pm 1.23}$	$\boldsymbol{94.47\pm 0.45}$	$\boldsymbol{94.27\pm 0.38}$	$\boldsymbol{94.79\pm 0.23}$	$\boldsymbol{92.62\pm 0.89}$	$\boldsymbol{96.32\pm 0.60}$
Fine-tuning	$70.02\pm 0.38$	$75.58\pm 1.12$	$71.31\pm 0.69$	$68.54\pm 0.49$	$82.10\pm 1.36$	$65.94\pm 0.33$
LwF (2016)	$79.06\pm 0.60$	$81.53\pm 0.34$	$79.90\pm 1.01$	$77.21\pm 0.94$	$77.63\pm 0.84$	$74.72\pm 0.79$
SI+ER (2017)	$81.38\pm 0.95$	$84.89\pm 0.52$	$83.14\pm 0.23$	$80.27\pm 0.48$	$92.17\pm 0.51$	$84.30\pm 0.87$

TABLE VI: Widar3.0 AFA and AFR (mean

\pm

sample SD, seeds 42–44) in further user orders and rooms.

\Delta

: paired AFA difference to the strongest listed baseline. Bold: best of each column

Without replay
Method	3DO		WiMANS
	AFA $\uparrow$	AFR $\downarrow$	AFA $\uparrow$	AFR $\downarrow$
WiMoT	$\boldsymbol{93.71\pm 2.19}$	$\boldsymbol{3.89\pm 3.26}$	$\boldsymbol{88.80\pm 0.89}$	$\boldsymbol{-0.31\pm 0.46}$
Fine-tuning	$84.56\pm 0.73$	$16.04\pm 1.30$	$86.51\pm 0.57$	$4.76\pm 1.16$
EWC	$84.56\pm 0.74$	$16.05\pm 1.32$	$86.52\pm 0.56$	$4.73\pm 1.13$
SI	$87.74\pm 1.19$	$10.59\pm 0.93$	$87.55\pm 0.28$	$3.40\pm 1.00$

TABLE VII: Robustness to domain order on 3DO and WiMANS (mean

\pm

SD over six orders; AFA in %, AFR in pp). Bold: best of each column; WiMoT is also bold when it ranks second

Panel of Figure 10 — Fig. 4: AFA lead of WiMoT over the strongest baseline of each setting against the AFR of that baseline in the 42 settings of RQ3 (domain orders, rooms, label budgets, and frozen encoders; three-seed means, Supplementary Table S7). Pearson $r=0.94$ , Spearman $\rho=0.94$ ; $r=0.80$ after the dataset means are removed

Encoder	3DO	WiMANS	Widar3.0
MantisV2 (pretrained)	94.02 ( $+3.77$ )	89.61 ( $+1.03$ )	95.16 ( $+13.78$ )
MantisV2 (random init.)	96.03 ( $+7.89$ )	78.29 ( $+1.44$ )	87.33 ( $+30.99$ )
MOMENT-1-large	94.82 ( $+7.30$ )	78.50 ( $-5.04$ )	86.20 ( $+16.62$ )
TSPulse-r1	95.56 ( $+9.56$ )	68.42 ( $-4.53$ )	84.69 ( $+26.40$ )
Chronos-2	93.88 ( $+6.31$ )	86.11 ( $-3.40$ )	88.12 ( $+13.70$ )
TimesFM 2.5	91.18 ( $+4.15$ )	81.58 ( $-2.67$ )	87.36 ( $+17.90$ )

TABLE VIII: AFA of WiMoT (%) with other frozen encoders and paired difference from the strongest of five baselines (pp). Bold: WiMoT has the highest AFA among itself and the five baselines

Panel of Figure 12 — Fig. 5: AFA of WiMoT against the SupCon weight (upper row, later domains, $\lambda_{\mathrm{h}}=1$ ) and the historical-loss weight (lower row, $\lambda_{\mathrm{sc}}=0.5$ ); mean $\pm$ SD over seeds 42–44. Star: weight of the paper; dashed line: strongest baseline of Table I ; open circle: weight 0. Vertical axes differ between datasets. Supplementary Table S12 lists all values

Panel of Figure 13 — Fig. 6: Deployment design. Queries need neither a domain identity nor stored samples. A new domain is encoded once, adds its moments to the store, and updates the adapter on cached basis features; its recordings are then deleted.

Quantity	3DO	WiMANS	Widar3.0
Inference per query (GFLOPs)	8.55	44.31	88.60
Arrival encoding per domain (TFLOPs)	20.5	16.7	39.5–39.9
Gradient updates per domain (TFLOPs) †	0.14	0.02	0.07
Prepared recordings per domain (MiB)	167.1	116.8	275.6–278.1
Aggregates added per domain (MiB)	7.0	19.1	13.0
Class moments, full storage (MiB)	18.04	54.11	108.21

TABLE IX: Computation and storage of WiMoT

Data	Order	WiMoT	ConSense	EdgeCL	Best classic
3DO	123	94.02	76.31	89.68	90.25 (SI+ER)
	132	94.03	80.21	86.05	88.99 (SI+ER)
	213	96.54	74.39	88.69	89.73 (EWC+ER)
	231	95.31	76.62	86.28	88.94 (SI+ER)
	312	90.64	78.19	83.88	89.51 (LwF+ER)
	321	91.75	78.06	83.99	89.44 (DER++)

TABLE S1: AFA (%) for each domain order; three-run means. Domains are numbered by their position in the original order; best classic: highest of the twelve classic baselines in that order

User	5	10	11	12	13	14	15	16	17
WiMoT	96.9	96.9	95.3	93.8	98.0	97.8	96.0	95.8	86.0
LwF+ER	80.0	75.3	81.1	82.9	83.6	86.9	82.0	90.9	75.8
SI+ER	85.8	78.0	78.0	84.0	81.1	84.7	78.9	80.9	81.1
CN-DPM	82.2	67.1	71.6	75.8	74.2	72.9	68.9	80.7	80.2
EdgeCL	73.3	61.3	73.6	81.6	76.2	74.4	73.3	81.3	80.0
ConSense	67.8	48.0	58.0	67.6	68.2	67.1	62.2	62.4	80.2

TABLE S2: Widar3.0 final test accuracy (%) per user after the ninth stage: three-run means

Without replay
Method	3DO		WiMANS		Widar3.0
Later-domain label fraction	25%	10%	25%	10%	25%	10%
Labels per class and later domain	200	80	11	5	19	8
WiMoT	$\boldsymbol{93.24\pm 0.73}$	$\boldsymbol{92.04\pm 0.21}$	$85.39\pm 0.94$	$82.72\pm 2.16$	$\boldsymbol{85.56\pm 1.02}$	$\boldsymbol{75.75\pm 1.15}$
Fine-tuning	$84.19\pm 1.44$	$83.11\pm 1.40$	$80.25\pm 2.83$	$75.62\pm 0.93$	$64.22\pm 0.49$	$57.80\pm 1.79$
SI (2017)	$87.90\pm 0.60$	$83.94\pm 1.88$	$80.66\pm 2.80$	$75.93\pm 0.82$	$65.60\pm 0.99$	$59.48\pm 1.58$

TABLE S3: AFA (%) with reduced labels in every later domain (mean

\pm

sample SD, three runs); the first domain is fully labeled, and the best value of each column is in bold

	3DO		WiMANS		Widar3.0
Encoder	WiMoT	Strongest baseline	WiMoT	Strongest baseline	WiMoT	Strongest baseline
MantisV2 (pretrained)	$\boldsymbol{94.02\pm 0.07}$	$90.25$ SI+ER ( $+3.77$ , 3/3)	$\boldsymbol{89.61\pm 0.47}$	$88.58$ LwF ( $+1.03$ , 3/3)	$\boldsymbol{95.16\pm 1.23}$	$81.38$ SI+ER ( $+13.78$ , 3/3)
MantisV2 (random init.)	$\boldsymbol{96.03\pm 0.46}$	$88.14$ EdgeCL ( $+7.89$ , 3/3)	$\boldsymbol{78.29\pm 1.39}$	$76.85$ EdgeCL ( $+1.44$ , 3/3)	$\boldsymbol{87.33\pm 1.08}$	$56.35$ SI+ER ( $+30.99$ , 3/3)
MOMENT-1-large	$\boldsymbol{94.82\pm 1.22}$	$87.53$ DER++ ( $+7.30$ , 3/3)	$78.50\pm 0.99$	$83.54$ SI+ER ( $-5.04$ , 0/3)	$\boldsymbol{86.20\pm 0.73}$	$69.58$ LwF ( $+16.62$ , 3/3)
TSPulse-r1	$\boldsymbol{95.56\pm 0.71}$	$86.01$ DER++ ( $+9.56$ , 3/3)	$68.42\pm 0.78$	$72.94$ DER++ ( $-4.53$ , 0/3)	$\boldsymbol{84.69\pm 0.30}$	$58.30$ LwF ( $+26.40$ , 3/3)
Chronos-2	$\boldsymbol{93.88\pm 0.39}$	$87.57$ DER++ ( $+6.31$ , 3/3)	$86.11\pm 1.35$	$89.51$ SI+ER ( $-3.40$ , 0/3)	$\boldsymbol{88.12\pm 0.31}$	$74.42$ LwF ( $+13.70$ , 3/3)

TABLE S4: AFA (%) with other frozen encoders (mean

\pm

sample SD, three runs), the strongest of five baselines (fine-tuning, LwF, SI+ER, DER++, EdgeCL), and the paired AFA difference of WiMoT from it (pp; number of seeds with higher AFA). Bold: WiMoT is higher than the strongest baseline

Encoder	Param.	$D$	Revision	Weights	PCA variance (%)
MOMENT-1-large	341.2	1024	ca58581b	a56928052ac6	99.11 / 98.75 / 98.87
TSPulse-r1	1.08	3312	2e64fcdc	57fa03b67d14	97.95 / 99.41 / 99.26
Chronos-2	119.5	768	29ec3766	ddcda3c7508b	99.86 / 99.93 / 99.90
TimesFM 2.5 (200M)	231.3	1280	1d952420	2f776efe6245	99.17 / 99.45 / 99.38
TimesFM 3.0	330.7	1280	43046b85	a7592b0a8432	99.00 / 99.37 / 99.32
MantisV2, random init.	4.19	512	–	–	no PCA

TABLE S5: Frozen encoders: parameters of the loaded checkpoint (millions), per-channel embedding width

D

, repository revision, weight digest (first 12 hex digits), and variance retained by the 512-coordinate PCA (3DO / WiMANS / Widar3.0)

	3DO		WiMANS		Widar3.0
Storage	MiB	$\Delta$ AFA	MiB	$\Delta$ AFA	MiB	$\Delta$ AFA
Full	18.04	$94.03$ %	54.11	$89.61$ %	108.21	$95.19$ %
Upper triangle	9.05	identical	27.16	identical	54.32	identical
Rank 128	4.54	$+0.31\pm 0.49$ (2/1)	13.63	identical	27.26	identical
Rank 32	1.16	$+1.07\pm 0.62$ (3/0)	3.49	$+0.00\pm 0.62$ (1/1)	6.97	$-0.49\pm 0.31$ (0/3)
Rank 8	0.32	$+0.25\pm 0.79$ (2/1)	0.95	$-0.41\pm 0.78$ (1/2)	1.90	$-2.74\pm 0.39$ (0/3)

TABLE S6: Covariance storage: class-moment payload (MiB) and paired AFA difference from full storage (pp, mean

\pm

SD over seeds 42–44; number of seeds higher/lower)

3DO				WiMANS				Widar3.0
Setting	Lead	AFR best	AFR mean	Setting	Lead	AFR best	AFR mean	Setting	Lead	AFR best	AFR mean
Main	$+3.77$	$5.64$	$7.27$	Main	$+0.41$	$2.78$	$4.44$	Main	$+13.11$	$10.25$	$16.21$
Order 1-3-2	$+5.04$	$7.43$	$9.12$	Order 1-3-2	$-1.03$	$0.62$	$3.69$	Order 4-8-9-2-3-5-1-7-6	$+11.14$	$11.67$	$17.26$
Order 2-1-3	$+6.81$	$7.24$	$9.83$	Order 2-1-3	$+1.85$	$1.85$	$3.78$	Order 6-7-4-3-1-2-5-8-9	$+14.52$	$14.33$	$19.31$
Order 2-3-1	$+6.36$	$6.54$	$9.18$	Order 2-3-1	$+1.03$	$2.93$	$3.08$	Reverse order	$+9.58$	$10.33$	$15.08$
Order 3-1-2	$+1.13$	$4.19$	$9.42$	Order 3-1-2	$-1.95$	$-0.31$	$1.81$	Room 2	$+0.45$	$3.80$	$8.59$

TABLE S7: AFA lead of WiMoT over the strongest baseline of each setting (pp), AFR of that baseline (pp), and mean AFR of all baselines of the setting (pp); three-seed means

	3DO			WiMANS			Widar3.0
Method	9216	9728	512	9216	9728	512	9216	9728	512
WiMoT (Table I)	–	$94.02\pm 0.07$	–	–	$89.61\pm 0.47$	–	–	$95.16\pm 1.23$	–
WiMoT, amplitude stem b	–	–	$97.37\pm 0.23$	–	–	$25.82\pm 1.25$	–	–	$85.90\pm 0.37$
WiMoT, amplitude basis a	–	–	$99.02\pm 0.16$	–	–	$32.51\pm 2.06$	–	–	$94.00\pm 0.15$
SI+ER	$90.24\pm 0.22$	$90.25\pm 0.67$	$97.86\pm 0.59$	$87.96\pm 0.31$	$87.86\pm 0.64$	$23.97\pm 1.39$	$81.73\pm 0.62$	$81.38\pm 0.95$	$52.54\pm 1.37$
DER++	$89.76\pm 0.78$	$89.81\pm 0.89$	$95.89\pm 1.04$	$88.27\pm 0.53$	$88.27\pm 0.53$	$24.90\pm 1.39$	$80.27\pm 0.04$	$80.25\pm 1.01$	$46.79\pm 1.30$

TABLE S8: Test AFA (%, mean

\pm

sample SD, seeds 42–44) with the Mantis block (9216), the concatenation of the Mantis block and the amplitude descriptor (9728), and the amplitude descriptor alone (512)

	3DO		WiMANS		Widar3.0
Method	9216	9728	9216	9728	9216	9728
WiMoT	–	$\boldsymbol{94.02\pm 0.07}$	–	$\boldsymbol{89.61\pm 0.47}$	–	$\boldsymbol{95.16\pm 1.23}$
Fine-tuning	$82.98\pm 0.66$	$84.19\pm 1.13$	$86.52\pm 0.18$	$86.93\pm 1.17$	$72.47\pm 1.57$	$70.02\pm 0.38$
EWC	$82.97\pm 0.61$	$84.19\pm 1.10$	$86.32\pm 0.18$	$86.93\pm 1.17$	$72.47\pm 1.24$	$70.62\pm 0.37$
SI	$89.69\pm 0.55$	$89.38\pm 0.21$	$87.14\pm 0.71$	$87.76\pm 1.52$	$77.53\pm 0.69$	$75.51\pm 0.28$
MAS	$87.55\pm 0.69$	$87.72\pm 1.13$	$86.93\pm 0.18$	$88.17\pm 1.28$	$68.77\pm 0.31$	$68.44\pm 0.58$

TABLE S9: Test AFA (%, mean

\pm

sample SD, seeds 42–44) of the twenty-one baselines with the 9216-dimensional Mantis block alone and with the 9728-dimensional input (Table I), and the paired AFA difference of WiMoT from the strongest baseline (pp; seeds ahead). Bold: highest baseline of each column. The last two rows compare with the better of the two variants of every baseline, listed under 9728

	3DO			WiMANS			Widar3.0
Method	New	Earlier	AFR	New	Earlier	AFR	New	Earlier	AFR
WiMoT (full)	93.10	96.75	0.26	87.22	90.83	$-0.28$	94.97	96.38	0.42
w/o SupCon	87.50	97.22	$-0.57$	80.37	85.37	$-3.89$	88.52	94.02	0.15
w/o $\mathcal{L}_{\mathrm{hist}}$	93.09	96.58	0.62	86.94	89.35	0.83	94.93	95.67	1.10
w/o historical moments	97.92	66.73	30.54	86.48	82.96	5.00	94.47	43.65	52.45
Frozen adapter ( $t>1$ )	92.67	95.37	1.33	83.70	88.06	0.09	88.07	90.88	0.40

TABLE S10: New-domain accuracy and earlier-domain accuracy (%) and AFR (pp), as in Table III of the main paper, for the single-component ablation of Table IV; five runs, seeds 42–46

3DO
Classifier	AFA	AFR	New	Earlier
OSCL (WiMoT)	93.98 $\pm$ 0.15	0.26 $\pm$ 0.31	93.10 $\pm$ 0.13	96.75 $\pm$ 0.34
Nearest class mean	80.89 $\pm$ 0.89	14.04 $\pm$ 0.98	88.91 $\pm$ 0.97	78.95 $\pm$ 0.41
LDA	92.06 $\pm$ 0.49	1.64 $\pm$ 0.90	90.98 $\pm$ 0.89	95.69 $\pm$ 0.63
QDA	91.00 $\pm$ 0.70	1.23 $\pm$ 0.26	89.53 $\pm$ 0.98	95.47 $\pm$ 0.23
Linear probe $W^{\dagger}$	85.77 $\pm$ 0.49	$-$ 9.97 $\pm$ 0.48	70.80 $\pm$ 0.97	98.22 $\pm$ 0.35

TABLE S11: AFA, AFR (pp), and new- and earlier-domain accuracy (%) with the OSCL classifier replaced by other classifiers fitted on the same adapters and transported moments; mean

\pm

SD, seeds 42–46. The OSCL rows are the full-WiMoT rows of Table IV; ridge rows use a fixed

\beta=10^{-3}

. † Fitted to earlier domains only (Section S-VIII )

SupCon weight $\lambda_{\mathrm{sc}}$ (later domains), $\lambda_{\mathrm{h}}=1$
	3DO				WiMANS				Widar3.0
Weight	AFA	AFR	New	Earlier	AFA	AFR	New	Earlier	AFA	AFR	New	Earlier
0	94.11 $\pm$ 0.47	$-$ 0.10 $\pm$ 0.74	92.96 $\pm$ 0.76	96.92 $\pm$ 0.42	88.68 $\pm$ 0.47	$-$ 0.93 $\pm$ 0.46	85.65 $\pm$ 1.67	89.97 $\pm$ 0.27	93.80 $\pm$ 0.54	$-$ 0.14 $\pm$ 0.21	90.19 $\pm$ 0.59	95.42 $\pm$ 0.38
0.1	93.95 $\pm$ 0.27	0.10 $\pm$ 0.10	92.81 $\pm$ 0.58	96.92 $\pm$ 0.36	89.09 $\pm$ 1.17	$-$ 0.62 $\pm$ 0.96	86.57 $\pm$ 1.22	90.28 $\pm$ 0.80	94.96 $\pm$ 0.71	0.42 $\pm$ 0.25	94.56 $\pm$ 0.25	96.14 $\pm$ 0.82
0.25	93.94 $\pm$ 0.24	0.22 $\pm$ 0.24	92.90 $\pm$ 0.43	96.82 $\pm$ 0.39	89.30 $\pm$ 0.71	0.15 $\pm$ 0.27	87.19 $\pm$ 1.49	90.28 $\pm$ 0.46	95.21 $\pm$ 0.72	0.31 $\pm$ 0.19	94.86 $\pm$ 0.32	96.36 $\pm$ 0.84
0.5	94.02 $\pm$ 0.07	0.32 $\pm$ 0.33	93.07 $\pm$ 0.10	96.82 $\pm$ 0.46	89.61 $\pm$ 0.47	0.00 $\pm$ 0.46	87.50 $\pm$ 1.22	90.74 $\pm$ 0.00	95.16 $\pm$ 1.23	0.58 $\pm$ 0.58	95.08 $\pm$ 0.42	96.31 $\pm$ 1.13

TABLE S12: AFA, AFR (pp), and new- and earlier-domain accuracy (%) of WiMoT for different loss weights; mean

\pm

SD, seeds 42–44. Bold: weights of the paper. In the last block

\lambda_{\mathrm{sc}}

also acts on the first domain

Continual Learning

Continual Learning studies how models acquire knowledge from changing data streams while retaining or selectively revising what they learned before.

See all 301 papers

May 2, 2026

MU-SHOT-Fi: Self-Supervised Multi-User Wi-Fi Sensing with Source-free Unsupervised Domain Adaptation

Deep learning has been widely adopted for WiFi CSI-based human activity recognition (HAR) due to its ability to learn spatio-temporal features in a privacy-preserving and cost-effective manner. However, DL-based models generalize poorly across environments, a challenge amplified in multi-user settings where overlapping activities cause CSI entanglement and domain shifts. Practical deployments often limit access to labeled source data due to privacy constraints, motivating source-free adaptation using only unlabeled target-domain CSI and a pre-trained source model. In this paper, we propose MU-SHOT-Fi, a source-free unsupervised domain adaptation framework for single- and multi-user Wi-Fi sensing. MU-SHOT-Fi employs permutation-invariant set prediction with Hungarian matching during source training, followed by frozen-classifier backbone adaptation in the target domain. To enable stable adaptation without labels, we introduce occupancy-weighted information maximization that prevents model collapse by focusing diversity regularization on likely-occupied slots while excluding the dominant class from marginal entropy. Additionally, we employ binary rotation prediction as spatial self-supervision that exploits CSI frequency-time structure to learn domain-invariant features. For single-user scenarios, we introduce SU-SHOT-Fi by replacing occupancy weighting with standard information maximization and incorporating contrastive predictive coding to exploit temporal consistency. Extensive experiments on the WiMANS and Widar 3.0 datasets across cross-environment, cross-frequency, cross-orientation, and combined domain shifts demonstrate that MU-SHOT-Fi effectively recovers multi-user exact-activity classification performance under large domain shifts while maintaining accurate occupancy estimation and preventing collapse toward dominant classes.

Ahmed Y. Radwan, Hina Tabassum

department of Electrical Engineering and Computer Science, York University, Toronto, ON M3J 1P3, Canada

Source-Free Domain Adaptation Human Activity Recognition Self-Supervised Learning Permutation-Invariant Learning Invariant Representation Learning Wi-Fi Sensing

Ask PDF

Sep 25, 2026

CSI-Agent: LLM-Assisted Few-Shot Adaptation for Cross-Domain Wi-Fi CSI Sensing

Wi-Fi channel state information (CSI) has enabled device-free sensing applications such as human activity recognition. However, CSI sensing models remain brittle in cross-domain deployment, where changes in users or environments can produce incorrect predictions. Existing solutions usually treat this problem as an offline model-design problem, by pretraining a stronger representation or applying one fixed adaptation method to the entire target domain. In practice, labeled target data are scarce and different classes may fail in different ways under the same domain shift. To address this, we propose CSI-Agent, an evidence-seeking LLM agent that reformulates cross-domain CSI adaptation as a deployment-time decision-making problem. Rather than processing raw CSI or making sample-level predictions, CSI-Agent summarizes target-domain behavior into sensing-grounded class-level evidence. It establishes a strong target-adaptive default from complementary CSI views and uses an LLM planner to determine whether each class should retain the default or invoke a specialized action. Deterministic verification and bounded execution further reduce unreliable interventions. We evaluate CSI-Agent on four public datasets using five cross-domain splits covering device, user, environment, and compositional shifts. Under 1-shot adaptation, CSI-Agent achieves the best target-domain performance across all splits and improves the average Macro-F1 by about 16% compared to the strongest baseline method.

Tianya Zhao, Chuan Liu, Xuyu Wang

Knight Foundation School of Computing and Information Sciences, Florida International University, Miami, FL 33199, US

Few-Shot Domain Adaptation Wi-Fi Sensing

Ask PDF

Dec 14, 2025

FuseFi: Combining Irregularly Sampled CSI from Diverse Communication Packets and Frequency Bands for Wi-Fi Sensing

Existing Wi-Fi sensing systems rely on injecting high-rate probing packets to extract channel state information (CSI), leading to communication degradation and limited deployment flexibility. Although Integrated Sensing and Communication (ISAC) is a promising direction, existing solutions still rely on auxiliary packet injection because they exploit only uniform CSI from a single frame type, discarding approximately 70% of naturally available packets. We present FuseFi, a novel Wi-Fi-based ISAC framework that directly exploits irregularly sampled CSI from diverse communication packets across multiple frequency bands, eliminating intrusive packet injection and introducing no sensing-specific communication overhead. FuseFi integrates a CSI sanitization pipeline to harmonize heterogeneous packets and remove burst-induced redundancy, together with a time-aware attention model that learns directly from non-uniform CSI sequences without resampling. We further introduce CommCSI-HAR, a new dataset with irregularly sampled CSI from real-world dual-band communication traffic. Extensive evaluations on this dataset and, separately, on four public sensing tasks across three benchmark datasets show that FuseFi achieves state-of-the-art accuracy with a compact model size, while fully preserving communication throughput.

Gaofeng Dong, Kang Yang, Mani Srivastava

Department of Electrical and Computer Engineering, University of California, Los Angeles, CA, USA

Integrated Sensing and Communication Wireless Communications Wi-Fi Sensing

Ask PDF

Abstract

Continual Learning

Explore similar work

MU-SHOT-Fi: Self-Supervised Multi-User Wi-Fi Sensing with Source-free Unsupervised Domain Adaptation

CSI-Agent: LLM-Assisted Few-Shot Adaptation for Cross-Domain Wi-Fi CSI Sensing

FuseFi: Combining Irregularly Sampled CSI from Diverse Communication Packets and Frequency Bands for Wi-Fi Sensing