FedFiTS: Multifaceted Federated Learning

• FedFiTS is a term describing diverse federated learning frameworks used in wearable fitness tracking, few-shot image classification, and healthcare scheduling.
• It emphasizes privacy-preserving on-device training, parameter-efficient model updates with FiLM adapters, and trust-aware client selection to optimize communication.
• Despite common themes in communication efficiency and personalization, FedFiTS remains ambiguous due to its multiple independent definitions and application-specific adaptations.

FedFiTS is not a single universally standardized federated learning construct. In recent arXiv usage, the name denotes at least three distinct systems: a production-oriented fitness-tracking system naturally grounded in the FedFitTech baseline for inertial-sensor human activity recognition, a federated version of FiLM Transfer with Statistics for parameter-efficient few-shot image classification, and a healthcare-oriented framework for Fitness-Selected, Slotted Client Scheduling. A related communication paper on FedFT further notes that queries for “FedFiTS” may sometimes refer instead to a frequency-space transmission method rather than to a separate algorithmic family (Oz et al., 20 Jun 2025, Shysheya et al., 2022, Kahenga et al., 23 Sep 2025, Palihawadana et al., 2024).

1. Nomenclature and scope

The principal usages of the term occupy different parts of the federated learning design space. One usage is application-specific and centers on wearable fitness tracking; one is model-architectural and emphasizes parameter-efficient adaptation; and one is systems-oriented, focusing on trustworthy client selection and robust aggregation in healthcare. The resulting ambiguity is substantive rather than terminological only: each usage changes what is communicated, how clients participate, and what notion of personalization is prioritized.

Usage of “FedFiTS”	Domain	Defining mechanism
Production system grounded in FedFitTech	Wearable FitTech / HAR	Flower-based FL with TinyHAR, FedAvg, and client-side early stopping
FiLM Transfer with Statistics in FL	Few-shot image classification	Frozen backbone, FiLM adapters, locally built Naive Bayes or ProtoNets heads
Fitness-Selected, Slotted Client Scheduling	Healthcare and cross-domain FL	Three-phase participation, dynamic fitness/trust scoring, slotted aggregation
Name occasionally conflated with FedFT	General FL communication	DCT-based frequency-space transmission of model updates

The ambiguity matters because “FedFiTS” can refer either to an end-to-end application blueprint, to a specific parameterization of federated personalization, or to a selective scheduling framework. The FedFT paper explicitly states that it introduces FedFT rather than a separate method called “FedFiTS,” and that such wording most likely refers to the same method only in the context of that query (Palihawadana et al., 2024).

2. FedFiTS as a FitTech system grounded in FedFitTech

In the FitTech usage, FedFiTS is best understood as a production system naturally derived from the FedFitTech baseline. FedFitTech is a federated learning baseline under the Flower framework for privacy-preserving, communication-efficient human activity recognition from wearable inertial sensors. It packages server orchestration, client-side data processing and training, global aggregation, and a case-study mechanism for client-side early stopping. The motivating setting is fitness tracking with wrist-worn accelerometers and gyroscopes, where centralized learning requires uploading raw sensor traces and thereby raises privacy, regulatory, and communication concerns. The baseline instead keeps raw data on-device and exchanges model parameters or updates (Oz et al., 20 Jun 2025).

The system architecture follows the canonical synchronous FL loop. The server coordinates rounds, broadcasts the current global model, receives updated local models, and aggregates them with FedAvg. The client role, implemented on a wearable or phone, is to collect inertial sensor data at 50 Hz, segment it into fixed windows of 100 samples, train locally for one epoch per round with Adam at learning rate 0.001 and batch size 32, compute local validation F1-scores for early stopping, and upload model parameters or gradients depending on the implementation. The baseline runs for 100 global rounds and assumes a homogeneous smartwatch-like environment in which most clients participate each round unless they early-stop or are unavailable. The data flow is sensor streams $\rightarrow$ windowing $\rightarrow$ optional feature extraction $\rightarrow$ local training $\rightarrow$ model update $\rightarrow$ server aggregation $\rightarrow$ redistributed global weights (Oz et al., 20 Jun 2025).

Federated optimization is expressed as minimization of a weighted sum of local objectives,

$$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$

with FedAvg aggregation

$$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$

TinyHAR is the local model. It combines per-channel convolutions, a transformer block with self-attention for cross-channel interactions, a fully connected fusion, an LSTM for temporal dependencies, and a final self-attention layer, with the stated goal of balancing accuracy and low parameter count for on-device training. Personalization enters through local training and through early stopping: a client ceases participation when its local validation F1-score stabilizes over a sliding window with patience $=5$ and threshold $\tau=0.01$. Communication cost is written as $\rightarrow$0, where $\rightarrow$1 is the number of rounds attended and $\rightarrow$2 is model size in bytes transferred per round (Oz et al., 20 Jun 2025).

The evaluation uses the WEAR inertial dataset with 24 clients, derived from 22 participants with repeated recordings treated as separate devices. The activity taxonomy comprises 18 outdoor fitness activities plus a NULL class: five jogging labels, five stretching labels, and eight strength labels. Sampling is at 50 Hz with 2-second windows, and the split allocates the first 20% of each label’s time series per client to test and the remaining 80% to train; a local validation subset supports early stopping. Reported findings are descriptive rather than inferential: 9 of 24 clients stopped early; the earliest stop was at round 40; overall redundant communications were reduced by 13%; baseline mean F1 across clients was 68% and early stopping mean F1 was 67%; 11 of 24 clients improved overall F1 under early stopping; Client 18 reduced its own communication by 47% and improved from 71% to 74% F1; and one cited per-label example is push-ups for Client 2 improving from 0% to 76% F1. This usage of FedFiTS therefore denotes a privacy-aware FitTech deployment pattern that seeks a balance between population-level generalization and device-specific nuances rather than a new FL optimizer in itself (Oz et al., 20 Jun 2025).

3. FedFiTS as federated FiLM Transfer with Statistics

A distinct usage appears in the FiT literature, where FedFiTS denotes “FiLM Transfer with Statistics” in the federated setting. Here the central problem is personalized and federated few-shot image classification under severe communication constraints. The architecture freezes a high-capacity pretrained backbone $\rightarrow$3, inserts small FiLM adapter layers with parameters $\rightarrow$4, and constructs a classifier head locally from support data rather than communicating a learned dense head. For an activation map $\rightarrow$5, the FiLM transformation is

$\rightarrow$6

In ResNetV2 blocks, FiLM follows the middle $\rightarrow$7 convolution, with an additional FiLM before the final head. The scalars are initialized to $\rightarrow$8, $\rightarrow$9, and with a FiLM layer in each block of ResNet50 the adapter set has 11,648 parameters, which is stated as less than 0.05% of the backbone (Shysheya et al., 2022).

The classifier head replaces the standard linear layer with an automatically configured Gaussian Naive Bayes head $\rightarrow$0. For class $\rightarrow$1, the model estimates $\rightarrow$2, $\rightarrow$3, and $\rightarrow$4 from support features $\rightarrow$5 and predicts with

$\rightarrow$6

The means and covariances are estimated by maximum likelihood on the support set. Three covariance models are used: QDA, LDA, and ProtoNets. In LDA, shared covariance yields linear discriminants, and one useful implementation stores for each class the vector $\rightarrow$7 and scalar $\rightarrow$8, computing logits as $\rightarrow$9. The paper’s training procedure is episodic: at each iteration, it samples support and query sets, recomputes classifier statistics from support data, and optimizes $\rightarrow$0 and $\rightarrow$1 by maximizing query likelihood with Adam at learning rate 0.0035 over 400 iterations, with support set size 100 and no data augmentation. For 1-shot, training is skipped and the initialization $\rightarrow$2 and $\rightarrow$3 is used (Shysheya et al., 2022).

In the federated variant, only FiLM adapter parameters $\rightarrow$4 are communicated between server and clients. The Naive Bayes or ProtoNets head is constructed locally from each client’s support data and is not sent. Server aggregation uses FedAvg on adapter parameters,

$\rightarrow$5

where $\rightarrow$6 denotes $\rightarrow$7 on the server. This partitioning is the defining trait of the FedFiTS design in this line of work: personalization is achieved by locally recomputed class statistics and optional local FiLM fine-tuning, while global communication is restricted to a very small parameter subset (Shysheya et al., 2022).

Parameter efficiency is quantified explicitly. For a BiT-M-R50x1 backbone with output dimension $\rightarrow$8 and $\rightarrow$9 classes, the paper reports $\rightarrow$0 for full BiT fine-tuning, approximately $\rightarrow$1 for FiT-QDA, $\rightarrow$2 for FiT-LDA, and $\rightarrow$3 for FiT-ProtoNets. The corresponding update ratio for FiT-LDA is

$\rightarrow$4

which is reported as less than 0.14%. In CIFAR100 federated experiments over 60 rounds with 5 clients per round, FiT transmits 0.1M parameters per round and 7M overall, compared with 237M per round and 14B overall for BiT, which the paper characterizes as more than three orders of magnitude less communication. Empirically, the method achieves comparable accuracy to BiT in global FL across varying client counts and shots, and strong personalized performance by initializing from aggregated FiLM parameters and constructing local ProtoNets heads. This usage of FedFiTS is therefore a communication-efficient personalization scheme rather than a client-scheduling policy (Shysheya et al., 2022).

4. FedFiTS as Fitness-Selected, Slotted Client Scheduling

A third and later usage names a healthcare-oriented FL framework: “Fitness-Selected, Slotted Client Scheduling.” In this formulation, FedFiTS addresses non-IID data, heterogeneous and unreliable clients, adversarial manipulation, and fairness requirements. The framework combines fitness-based client election, slotted participation, dynamic thresholding, trust-aware aggregation, and cohort-based scheduling. Its motivation is that conventional FedAvg can amplify statistical bias, admit poisoned updates, and waste rounds under drifting client utility when clients are sampled randomly and aggregated uniformly (Kahenga et al., 23 Sep 2025).

The core client score combines quality of learning and data quantity. Let $\rightarrow$5 be the global model at round $\rightarrow$6 and $\rightarrow$7 the local model for client $\rightarrow$8 at round $\rightarrow$9. On client $\rightarrow$0, the global model and the local model are evaluated on the local test set to obtain $\rightarrow$1 and $\rightarrow$2, respectively. Define the midpoint in the loss-accuracy plane,

$\rightarrow$3

and let $\rightarrow$4 be the loss unit vector. The quality measure is an angle

$\rightarrow$5

where smaller angle is interpreted as better quality of learning. Data quality fraction is $\rightarrow$6 with $\rightarrow$7. The round score is

$\rightarrow$8

Rather than fixing $\rightarrow$9, the framework computes a dynamic $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$0 by averaging client-wise maximizers, yielding $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$1 when many clients have $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$2 and $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$3 otherwise. Client election is based on an adaptive threshold,

$$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$4

with $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$5 controlling openness (Kahenga et al., 23 Sep 2025).

FedFiTS then applies a three-phase participation strategy. In free-for-all training, all available clients train during initial rounds to bootstrap score estimates. In natural selection, the server computes scores and elects clients satisfying $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$6. In slotted team participation, the selected set remains active across a slot of consecutive rounds, regulated by a maximum slot length MSL and a performance fluctuation threshold PFT. Slot stability is tracked with a slot-level quantity $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$7 and a counter of consecutive declines,

$$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$8

Reselection is triggered when $$F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w), \qquad F_k(w)=\mathbb{E}_{\xi\sim D_k}[\ell(w;\xi)],$$9 or when $$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$0. Trust can augment fitness through penalties on cosine-similarity outliers, update-magnitude outliers, stale updates, and suspicion flags:

$$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$1

Aggregation uses trust or fitness weights $$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$2 and supports weighted averaging as well as robust fallbacks such as coordinate-wise median, trimmed mean, and Krum (Kahenga et al., 23 Sep 2025).

The paper also gives a convergence analysis. Under assumptions of $$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$3-smoothness, unbiased stochastic gradients with bounded variance, bounded heterogeneity, selection floors or bias control, bounded robust-aggregation error, bounded local or slot drift, and suitable stepsizes, it collects residuals into

$$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$4

It then states a one-step descent lemma and a non-convex stationarity bound,

$$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$5

as well as a linear convergence corollary under the Polyak–Łojasiewicz condition. Communication complexity is contrasted with FedAvg through

$$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$6

with additional lightweight selection overhead occurring only when reselection is invoked (Kahenga et al., 23 Sep 2025).

The experimental section covers Chest X-ray pneumonia, MNIST, Fashion-MNIST, and Crop Recommendation. On MNIST, using all clients per round with slot size 1, the paper reports, for example, that at $$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$7 under normal conditions FedAvg reaches accuracy 0.959 in 6257.6 seconds whereas FedFiTS reaches 0.989 in 5133.5 seconds; under attack, FedFiTS improves accuracy by 0.006 to 0.049 while also reducing time. On X-ray pneumonia at $$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$8, the reported accuracies are 0.968 for FedFiTS under normal conditions and 0.940 under attack, compared with 0.946/0.913 for FedRand and 0.950/0.913 for FedPow. Participation fairness is summarized through proxy metrics: participation ratio rises to 82% with dynamic $$w_{t+1}=\sum_{k\in S_t}\frac{n_k}{\sum_{j\in S_t}n_j}\,w_t^{(k)}.$$9 versus 45% for FedAvg and 52% for FedPow. In this usage, FedFiTS is a trust- and fairness-aware client orchestration method rather than an architectural adaptation scheme (Kahenga et al., 23 Sep 2025).

5. Relation to FedFT and frequency-space communication

A separate source introduces FedFT, not FedFiTS, but explicitly remarks that the queried name may be intended to refer to that method. FedFT stands for federated frequency-space transformation and reduces communication overhead by sending Discrete Cosine Transform coefficients instead of raw tensor-space model parameters. Two transmission variants are considered: the DCT of the full local model and the DCT of parameter deltas $=5$0. The paper states that communicating deltas in frequency space is more compact because model updates have lower variance than full weights, leading to tighter spectral energy concentration and making high-frequency coefficients easier to prune without materially harming reconstruction (Palihawadana et al., 2024).

The method uses DCT-IV because it is linear, orthogonal, and symmetric. For a $=5$1D tensor $=5$2, the transform and inverse transform are written explicitly, and linearity yields

$=5$3

This allows a FedAvg-like update directly in frequency space:

$=5$4

Clients receive $=5$5, reconstruct with IDCT, train locally, compute $=5$6, transform to $=5$7, prune coefficients by magnitude at rate $=5$8, and transmit the compressed coefficients. The reported per-client upstream communication reduction is roughly 5% to 30% depending on dataset and pruning rate, with runtime overhead up to about 6% relative to FedAvg or FedSim and at most 5% relative to FedProx. Example reductions at round 200 include MNIST from about 40 MB at $=5$9 to about 29 MB at $\tau=0.01$0, FEMNIST from about 104 MB to about 92.8 MB, Fed-MEx from about 45.6 MB to about 33.6 MB, and Fed-Goodreads from about 25.6 MB to about 13.8 MB. The paper therefore treats the “FedFiTS” wording as a naming confusion rather than as a formal synonym, but the confusion is understandable because both terms refer to communication-efficient FL mechanisms (Palihawadana et al., 2024).

6. Comparative interpretation and open issues

Taken together, the various FedFiTS usages describe three different levers for federated learning. The FitTech interpretation modifies the application pipeline around wearable HAR and emphasizes privacy-preserving local training plus early stopping. The FiLM Transfer interpretation modifies the parameterization of adaptation and the partition between communicated and local parameters. The healthcare scheduling interpretation modifies who participates, for how long, and how their updates are weighted. This suggests that “FedFiTS” functions in current literature as a family name only in a loose, non-standard sense, with meaning determined by domain and paper lineage rather than by a shared canonical algorithm (Oz et al., 20 Jun 2025, Shysheya et al., 2022, Kahenga et al., 23 Sep 2025).

Despite their differences, the variants converge on several recurring themes. One is communication efficiency: the FitTech baseline measures savings through reduced participation rounds, the FiT-based method through sub-1% updateable parameter ratios and local head construction, and the healthcare scheduler through restricting communication to selected cohorts rather than all clients. Another is personalization under heterogeneity: FedFitTech allows clients to disengage when global updates no longer improve local validation; FiT-based FedFiTS keeps local class statistics on-device and optionally fine-tunes only small adapters; the healthcare framework balances data quantity and learning quality through dynamic $\tau=0.01$1 and thresholding. A third shared theme is robustness. In FitTech, the challenge is user-specific activity execution and data imbalance; in FiT-based FedFiTS, it is few-shot overfitting and non-IID personalization; in the healthcare scheduler, it includes poisoning, staleness, and fairness-aware participation. A plausible implication is that the name’s recurring appeal comes from its association with selective information exchange under local constraints, even though the formal mechanisms differ.

The listed future directions also differ by lineage. In the FitTech baseline, open problems include label scarcity, stronger robustness to heterogeneity through methods such as SCAFFOLD, clustered FL, Meta-HAR, and ProtoHAR, tighter energy management through adaptive local epochs or layer freezing, integration of differential privacy and secure aggregation, multimodal fusion, and transfer learning for inertial representations (Oz et al., 20 Jun 2025). In the FiT-based line, extensions include alternative adapters such as LoRA, Adapters, VPT, AdaptFormer, Convpass, and CaSE, as well as compression or quantization of adapter updates and two-stage personalization strategies (Shysheya et al., 2022). In the healthcare scheduler, open issues include possible gaming or noise in the $\tau=0.01$2 metric, adaptive scheduling of $\tau=0.01$3, secure implementation of trust signals under secure aggregation, stronger guarantees against backdoors, group-level fairness metrics beyond participation, and tighter theoretical links between $\tau=0.01$4, MSL/PFT, and residual terms such as $\tau=0.01$5 and $\tau=0.01$6 (Kahenga et al., 23 Sep 2025).

A common misconception is therefore that FedFiTS denotes one settled FL algorithm. The literature summarized here does not support that view. Instead, the term designates multiple technically distinct constructs: a FitTech deployment blueprint, a parameter-efficient few-shot federated model, and a trust-aware selective scheduler, with occasional confusion with the separate FedFT communication method.