Federated QUIC Classification

• Federated QUIC classification is an approach to analyze encrypted traffic using distributed learning that keeps raw flows local.
• It employs diverse feature representations—including statistical metrics, CNN-based image encodings, and metric-learning embeddings—to bypass payload encryption.
• Techniques like client-side buffering and FedAvg aggregation stabilize model training amid non-IID data and time-varying traffic dynamics.

Federated QUIC classification denotes the family of methods that classify encrypted QUIC traffic while keeping raw traffic data at distributed clients or network domains, and coordinating learning or state across a server or other shared control plane. In the literature, the task is motivated by the fact that QUIC is encrypted and multiplexed, so payload inspection is unavailable and classifiers must instead rely on statistical flow features, packet size/direction/timing sequences, image-like encodings, or inferred application-layer structure; at the same time, deployment is constrained by privacy, label scarcity, non-IID client distributions, and time-varying traffic dynamics (Jozsa et al., 12 Sep 2025, Peng et al., 2021, Luxemburk et al., 18 Feb 2025, Zhang et al., 2023).

1. Scope, targets, and problem formulation

Federated QUIC classification sits at the intersection of encrypted traffic analysis and federated learning. In the most direct formulation, a central server coordinates training across multiple clients that retain local QUIC traces and exchange only model updates, with the goal of producing a global classifier without centralizing raw traffic (Peng et al., 2021, Jozsa et al., 12 Sep 2025). The target label space varies across studies: service classification over seven common QUIC services such as Discord, Facebook-Graph, Google-WWW, Instagram, Snapchat, Spotify, and YouTube (Jozsa et al., 12 Sep 2025); five Google QUIC-based services such as Google Drive, YouTube, Google Doc, Google Search, and Google Music (Peng et al., 2021); exact SNI domain recognition in encrypted QUIC traffic (Luxemburk et al., 18 Feb 2025); and estimation of the number of HTTP/3 request-response pairs or responses in a connection (Gahtan et al., 2024).

A central premise of the field is that QUIC classification remains possible despite encryption. One line of work uses statistical flow features and packet-level metadata from the first packets of a flow (Jozsa et al., 12 Sep 2025), another uses sampled packet sequences (Peng et al., 2021), another learns reusable embeddings from packet size, inter-packet time, and direction (Luxemburk et al., 18 Feb 2025), and another reconstructs HTTP request-response structure from observable timing, length, and direction information without decrypting payloads (Zhang et al., 2023). This directly counters the common misconception that QUIC encryption eliminates all useful passive observables.

The federated setting introduces two recurring constraints. First, privacy and organizational boundaries make centralized raw-traffic collection undesirable or infeasible (Jozsa et al., 12 Sep 2025, Peng et al., 2021). Second, real deployments are heterogeneous: different organizations or network prefixes observe different service mixes, which yields non-IID client data, and traffic volumes fluctuate over the day, which can make synchronous local training unstable when some clients have too little data in a round (Jozsa et al., 12 Sep 2025).

2. Observable feature spaces and representation strategies

The feature space for federated QUIC classification is not singular. In one operationally oriented formulation, each flow is represented by 283 input features derived from raw flow statistics and packet-level information for the first 30 packets, including per-direction packet features for client-to-server and server-to-client traffic and per-direction flow statistics such as mean, standard deviation, minimum, and maximum of packet sizes and inter-arrival times; all features are scaled to $[0,1]$ with min-max scaling (Jozsa et al., 12 Sep 2025). This representation is tailored to service classification under backbone-network conditions.

A different formulation operates on sampled subflows rather than fixed per-flow summary statistics. In the federated semi-supervised approach, each sampled packet uses packet length and relative arrival time, and direction is encoded by multiplying packet length by $+1$ or $-1$ for forward and backward traffic; two sampling schemes are used, simple sampling and incremental sampling, each producing multiple 45-packet subflows from each flow (Peng et al., 2021). This design reduces computational complexity and storage pressure while representing each traffic flow as multiple CNN inputs.

Metric-learning work replaces task-specific feature engineering with a universal embedding function

$\Phi: Flow \to \mathbb{R}^d,$

with the final model using $d=256$ (Luxemburk et al., 18 Feb 2025). The input is a fixed-length sequence of 30 packets, padded with zeros if shorter, and each packet is represented by packet size, inter-packet time, and direction. The architecture called 30pktTCNET_256 uses a CNN-based backbone, a compression neck, ArcFace loss, and nearest-neighbor retrieval in embedding space using cosine similarity and FAISS IndexFlatIP (Luxemburk et al., 18 Feb 2025). The paper explicitly notes that a shared pretrained embedding is a natural federated backbone and that federated clients could potentially share a common pretrained encoder while keeping their local databases or labels private.

Image-based representations provide another standardized interface. VisQUIC converts packet arrival time, packet length, and packet direction into RGB images by binning time into $M$ equispaced bins, packet lengths into $N$ bins, counting packets per $(\text{time bin}, \text{length bin})$, and using the red and green channels for server-to-client and client-to-server packets, respectively, with $M=N=32$ in the main setup (Gahtan et al., 2024). This representation is explicitly presented as suitable for CNN-based learning and is naturally compatible with distributed training because it can be computed locally.

Rule-based semantic extraction occupies a complementary position. A state-machine-based inference pipeline estimates HTTP request and response objects from packet direction, QUIC header type, packet length, packet arrival time, and packet order or position in the flow, and then derives connection-level outputs such as number of request-response pairs, connection duration, total request/response sizes, and multiplexing level (Zhang et al., 2023). This suggests a layered federated architecture in which local nodes do not share raw packets, but instead share semantically richer inferred features.

3. Federated learning architectures and training protocols

Two distinct federated paradigms appear in the literature. The earlier one is federated semi-supervised learning. In that framework, clients hold unlabeled traffic data, the central server holds a small labeled dataset, and the system proceeds in three stages: data preprocessing, federated pretraining, and central server retraining (Peng et al., 2021). During federated pretraining, clients perform unsupervised training on local unlabeled data and send model updates to the server; the server aggregates them with FedAvg. After convergence, the pretrained weights are copied into a retraining model, which the server fine-tunes on labeled data to produce the final classifier. The pretrained model uses MSELoss, the retrained model uses CrossEntropyLoss, and both use Adam (Peng et al., 2021).

The later operational formulation is fully supervised federated learning over autonomous clients. The CESNET-QUIC22 study partitions the data into 14 distinct subsets corresponding to different IP ranges or organizations, treats them as 14 autonomous federated clients, and evaluates five federated optimizers in Flower/PyTorch: FedAvg, FedProx, FedAdam, FedAdagrad, and FedYogi (Jozsa et al., 12 Sep 2025). Training proceeds in 3-hour intervals over Weeks 46 and 47, yielding 112 FL rounds, with 10 local epochs per round, cross-entropy loss, Adam optimizer, learning rate 0.001, and batch size 64 for federated clients; a centralized baseline uses learning rate 0.01 and batch size 1024 (Jozsa et al., 12 Sep 2025). The validation set is used for local early stopping, and the test set is used to report each round’s performance.

Although not itself a federated learning paper, the embedding-transfer study outlines a third architecture that is directly relevant to federated deployment: a frozen shared encoder with local nearest-neighbor classification (Luxemburk et al., 18 Feb 2025). In that design, the expensive representation learning can be centralized once, after which each site embeds its local data and performs 1-NN or related retrieval locally. Because the paper explicitly argues that each site can maintain its own local database of embeddings for local traffic classes, this architecture reduces the need for end-to-end supervised retraining at each site.

Across these systems, FedAvg remains structurally central. The operational CESNET-QUIC22 study treats FedAvg as the main practical baseline and reports that it is the most stable under buffering, while the semi-supervised study uses FedAvg for server-side aggregation of client updates weighted according to local sample size (Jozsa et al., 12 Sep 2025, Peng et al., 2021).

4. Non-IID heterogeneity, temporal volatility, and stabilization

A major conceptual development in federated QUIC classification is the separation of statistical heterogeneity from temporal volatility. Statistical non-IID client data arise because different organizations or network prefixes observe different service mixes, so each client’s local distribution differs (Jozsa et al., 12 Sep 2025). Earlier federated work also highlighted a different challenge: obtaining labeled data is difficult in reality, which motivates a semi-supervised design in which unlabeled data are used on clients and labeled data are used on the server (Peng et al., 2021).

The operationally novel contribution is the recognition that temporal traffic volatility is a separate systems problem. Traffic volume changes dramatically over the day, and during low-traffic periods some clients have too little data to produce reliable gradients; in synchronous federated learning, weak or noisy local updates from these clients can destabilize the global model (Jozsa et al., 12 Sep 2025). The authors explicitly argue that prior federated traffic-classification work mostly addresses non-IID statistical heterogeneity, but not the operational instability caused by time-varying traffic volume.

The proposed remedy is a client-side FIFO buffer that decouples local training from immediate traffic arrivals. Instead of training on whatever data arrive in a 3-hour interval, each client accumulates flow records into a fixed-size buffer, starts a local training round only when enough samples have been collected, and discards the oldest records when the buffer overflows (Jozsa et al., 12 Sep 2025). In the concrete setup, each client maintains three fixed-size FIFO buffers: a training buffer with 6400 flows, a validation buffer with 914 flows, and a test buffer with 1828 flows, corresponding to a 70/10/20 split; newly arriving flows are partitioned according to this ratio and inserted into the appropriate buffer (Jozsa et al., 12 Sep 2025).

The claimed effect is operational rather than algorithmically elaborate. Buffering smooths local dataset size from round to round, prevents low-traffic intervals from producing tiny noisy minibatches, makes local updates more statistically meaningful, and renders synchronous global aggregation less sensitive to daily traffic dips (Jozsa et al., 12 Sep 2025). A common misconception is that more complex adaptive federated optimizers should dominate simple averaging in difficult environments; in this particular setting, the paper reports the opposite, namely that FedAvg performs best in terms of both average F1 and stability once buffering is introduced.

5. Benchmarks, datasets, and reported performance

The empirical literature spans several datasets and task formulations. CESNET-QUIC22 is used for both the operational federated service-classification study and the embedding-pretraining study (Jozsa et al., 12 Sep 2025, Luxemburk et al., 18 Feb 2025). The former uses Weeks 46 and 47 because the first two weeks exhibited stronger drift, and partitions the data into 14 autonomous clients (Jozsa et al., 12 Sep 2025). The latter uses the third week of 2022, W-2022-46, with 33.7 million samples drawn from a larger corpus of 153 million anonymized flows, and pretrains on 2000 remapped SNI domains with disjoint train, validation, and test domain splits (Luxemburk et al., 18 Feb 2025). A different federated semi-supervised study uses a public QUIC dataset captured at UC Davis, split into a pretrained dataset of 6,439 flows and a retrained dataset of 150 flows, with 100 clients and 80% of clients selected for training (Peng et al., 2021). VisQUIC contributes a large-scale benchmarking corpus of over 100,000 labeled QUIC traces from more than 44,000 websites collected over four months, with corresponding SSL keys and derived images (Gahtan et al., 2024).

Setting	Dataset/task	Reported result
Buffered synchronous FL	CESNET-QUIC22, seven-service classification	95.2% F1 with FedAvg + buffering
Centralized upper bound	CESNET-QUIC22, same task	97.5% F1
Federated semi-supervised learning	UC Davis QUIC dataset	91.08% accuracy with simple sampling; 97.81% with incremental sampling
QUIC domain-recognition pretraining	CESNET-QUIC22, held-out domains	94.83% top-1 accuracy; 79.35% macro recall
HTTP/3 response estimation	VisQUIC, known-server setting	92.6% CAP within $\pm 3$ for $+1$0

The most explicit federated systems comparison appears in the volatility study. The centralized non-private model reaches 97.5% F1, while buffered federated learning with FedAvg reaches 95.2% F1, only 2.3 percentage points below the centralized baseline (Jozsa et al., 12 Sep 2025). Without buffering, performance drops sharply during low-traffic nighttime periods, some clients—especially clients 10 and 14—fail to converge, and all five aggregators are affected (Jozsa et al., 12 Sep 2025). The paper further notes that all classes in the centralized baseline achieve more than 95% precision, with particularly strong performance on Discord and Spotify, while social media classes such as Instagram and Facebook-Graph are somewhat more confusable.

The federated semi-supervised results show a different trade-off. With simple sampling, accuracy is 91.08% in the federated setting versus 94.04% for centralized semi-supervised learning; with incremental sampling, the gap becomes minimal, at 97.81% versus 97.93% (Peng et al., 2021). The paper attributes the gain of incremental sampling to its ability to capture both short-range and long-range packet dependencies through progressively increasing sampling intervals.

The embedding-based literature broadens the notion of QUIC classification beyond direct federated training. The 30pktTCNET_256 model achieves 94.83% top-1 accuracy and 79.35% macro recall on disjoint test domains, while a raw input-space baseline reaches only 71.44% accuracy and 37.86% recall on the same pretraining task (Luxemburk et al., 18 Feb 2025). For downstream transfer, the pretrained embedding surpasses state of the art on four of five datasets, including UCDAVIS19 script, UTMOBILENET21, MIRAGE19, and MIRAGE22, but not CESNET-TLS22 (Luxemburk et al., 18 Feb 2025). The paper explicitly presents this as evidence that a reusable QUIC-trained representation can generalize across multiple encrypted-traffic tasks.

6. Adjacent distributed infrastructures, misconceptions, and open issues

Federated QUIC classification is not only a model-training problem; in some deployments it is also a distributed state-identification problem. QASM, a framework for QUIC-aware stateful middleboxes, is not presented as a federated learning protocol, but the paper states that it provides the closest relevant mechanism to federated or distributed QUIC classification through a Client Agent, a Tracking Agent, and QUIC-aware middleboxes that share a common logical view of a connection across IP and port changes (Selvam et al., 3 Feb 2026). The framework tracks $+1$1 O-DCID, $+1$2 Set of DCIDs, $+1$3 Set of client addresses, and $+1$4 Server IP and port, supports reactive query mode and proactive asynchronous push mode, and allows the Tracking Agent to be sharded based on DCID (Selvam et al., 3 Feb 2026). This matters because a classifier or middlebox policy tied only to the IP 5-tuple can be broken by QUIC connection migration.

Several misconceptions can therefore be stated precisely. First, QUIC encryption does not make traffic analysis impossible: service classification from 283 flow features, subflow-based CNN classification, domain-recognition embeddings, image-based HTTP/3 response estimation, and rule-based HTTP object inference all use observable metadata rather than payload inspection (Jozsa et al., 12 Sep 2025, Peng et al., 2021, Luxemburk et al., 18 Feb 2025, Gahtan et al., 2024, Zhang et al., 2023). Second, non-IID statistical heterogeneity is not the only deployment obstacle in federated learning; temporal volatility can destabilize synchronous training even when the statistical problem itself is well understood (Jozsa et al., 12 Sep 2025). Third, more complex federated optimizers do not necessarily outperform FedAvg in volatile real-traffic settings; the buffered CESNET-QUIC22 study reports FedAvg as the most stable and highest-performing tested method (Jozsa et al., 12 Sep 2025).

The open issues are correspondingly diverse. Some are explicit limitations of individual studies: evaluation restricted to Google services and relatively simple CNNs in the federated semi-supervised work (Peng et al., 2021); single-browser bias, page-request workload only, known-server evaluation, duplicate-image filtering, and the privacy sensitivity of SSL keys in VisQUIC (Gahtan et al., 2024); and the fact that the universal embedding paper discusses federated relevance as an implication rather than implementing a federated protocol directly (Luxemburk et al., 18 Feb 2025). Other issues concern data drift and implementation diversity. The CESNET-QUIC22 federated study excludes Weeks 44 and 45 because they exhibited stronger drift (Jozsa et al., 12 Sep 2025). A plausible implication is that QUIC implementation diversity and specification ambiguities documented in formal conformance testing—especially around migration, transport parameters, and error behavior—can act as an additional source of domain shift for classifiers trained on one stack and deployed on another (Crochet et al., 3 Mar 2025).

Taken together, the literature presents federated QUIC classification as a mature enough research area to support strong empirical results, but still one in which deployment realism matters. The strongest current conclusion is narrow and operational: privacy-preserving QUIC classification can approach centralized performance, but robust deployment requires architectures that account not only for non-IID data and scarce labels, but also for the daily ebb and flow of traffic and, in some networked settings, for distributed QUIC connection identity beyond the 5-tuple (Jozsa et al., 12 Sep 2025, Peng et al., 2021, Selvam et al., 3 Feb 2026).