💻 computer science

CAPE-Net: Conformal, Adaptive, and Drift-Aware Early-Exit Classification of Encrypted Traffic at ISP Scale

CAPE-Net is a conformal, adaptive, and drift-aware early-exit classifier that enables ISPs to perform accurate, certified, and maintainable encrypted traffic classification on packet-sequence metadata by integrating distribution-free error guarantees, abstention on unknown applications, and efficient per-class drift repair within a single compact model.

Original authors: Farzam Rezaei, Jorge E. López de Vergara, Luis de Pedro, Iván González

Published 2026-08-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Farzam Rezaei, Jorge E. López de Vergara, Luis de Pedro, Iván González

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Invisible Internet and the Race to Read the Unreadable

Imagine the internet as a massive, bustling city where data travels like cars on highways. For decades, the traffic police (Internet Service Providers, or ISPs) could easily see what each car was carrying. They could check the license plate (port numbers) or peek inside the trunk (payload inspection) to know if a vehicle was delivering pizza, carrying a package, or transporting a secret. But recently, the city changed. Almost every car now drives in a sealed, armored truck with tinted windows. This is the rise of encryption (like TLS and QUIC), which protects user privacy by scrambling the contents of the data so no one can read it.

This creates a tricky problem for the traffic police. They can't see the cargo anymore, but they still need to know what's inside to manage traffic jams, ensure emergency vehicles get through, or spot dangerous drivers. They can't open the trucks, but they can watch how the trucks move. They can see the size of the truck, which way it's turning, and how fast it accelerates or brakes. These are called "packet-sequence metadata." The big question for scientists is: Can we figure out what a truck is carrying just by watching its driving style? The answer is yes, but it's hard. The trucks are changing their driving habits, new types of vehicles are appearing, and the police need to make decisions instantly, not after the truck has already left the city.

CAPE-Net: The Smart Traffic Cop with a Crystal Ball

This paper introduces a new system called CAPE-Net (Conformal, Adaptive, and Drift-Aware Early-Exit Classification). Think of CAPE-Net not just as a traffic cop, but as a super-smart, multi-talented detective who can solve crimes before the suspect even leaves the room. The researchers built this system to handle the messy, real-world chaos of the internet, where traffic patterns change constantly and new apps appear out of nowhere.

Here is how CAPE-Net works, broken down into its three superpowers:

1. The "Early Exit" Strategy: Don't Wait for the Whole Movie
Most traffic classifiers are like movie critics who wait until the very last scene of a film to write their review. By then, the movie is over, and it's too late to change anything. CAPE-Net is different. It's like a critic who watches the first few minutes and says, "I know this is a comedy!" and stops watching. The system has 18 different "checkpoints" (like stopping points in a race). If the traffic flow looks easy to identify, CAPE-Net makes a decision after just 3 or 4 packets (tiny bits of data). If the flow is tricky, it waits a bit longer. On average, for standard internet traffic (TLS), it only needs to see about 7 to 8 packets to make a confident call, which is a huge speed boost compared to waiting for the whole flow.

2. The "Confidence Certificate": No Guessing Allowed
Usually, a computer might say, "I think this is a video stream," but it might be wrong. In the real world, a wrong guess can be expensive or dangerous. CAPE-Net uses a special math trick called conformal prediction to give itself a "confidence certificate." Before it makes a decision, it checks: "Am I sure enough?" If the answer is no, it refuses to guess. It says, "I don't know," rather than making a mistake. The paper shows that for about 98% of the different types of apps they tested, CAPE-Net can guarantee that when it does make a guess, it is right at least 95% of the time. It trades a little bit of speed for a lot of safety.

3. The "Drift Detector": Keeping Up with the Times
The internet is alive; it changes every day. Apps get updates, new features are added, and traffic patterns shift. A model trained last year might be confused today. CAPE-Net has a built-in alarm system. It watches its own performance like a teacher grading their own test. If it notices that a specific type of traffic is suddenly getting harder to identify (a "drift"), it flags it. Instead of rebuilding the whole system from scratch (which takes forever), it performs a tiny, surgical fix in just 16 to 19 seconds. It adjusts only the specific part of the brain dealing with that one tricky app, leaving the rest of the system untouched. This keeps the system running smoothly for weeks without needing a massive overhaul.

The Twist: What Happens with New Types of Trucks?
The researchers also tested CAPE-Net on a different kind of internet traffic called QUIC (used by many modern apps like YouTube and Google). QUIC trucks are bigger and take longer to get going. The system adapted perfectly, but it had to wait longer to make a decision—about 15 packets instead of 7 or 8. This proves that the system is flexible: it knows that different protocols need different amounts of evidence before it feels safe to speak up.

What CAPE-Net Says "No" To
The paper is very clear about what this system is not. It argues against the idea that you can just wait until a data flow is finished to classify it; that's too slow for real-time needs. It also argues against the idea that a single "one-size-fits-all" confidence threshold works for every app. Some apps are easy to spot, some are hard; CAPE-Net treats them differently. Finally, it rejects the idea that you need to retrain the entire massive model every time the internet changes. That's too expensive and disruptive. Instead, it suggests that small, targeted fixes are the way to go.

The Bottom Line
The authors found that by combining these three ideas—making decisions early, demanding high confidence before speaking, and fixing only what's broken—they created a system that is fast, reliable, and easy to maintain. They tested it on a massive dataset of 180 different applications and found that it works not just for today's traffic, but can adapt to tomorrow's changes with minimal effort. While the "quick fix" works great for standard internet traffic, the paper notes that for the newer, heavier QUIC traffic, the system might need a slightly different maintenance strategy. But overall, CAPE-Net shows that we can have our cake and eat it too: we can classify encrypted traffic quickly and accurately without needing to see the secret contents inside.

Technical Summary: CAPE-Net

Problem Statement

The near-universal adoption of encryption (TLS, Encrypted Client Hello, QUIC) has rendered traditional payload-based traffic classification obsolete. Network operators must now rely on packet-sequence metadata (sizes, directions, inter-arrival times) to classify traffic. However, existing solutions fail to meet the operational demands of a live ISP backbone, which requires a classifier to simultaneously:

Decide Early: Commit to a verdict on as few packets as possible to enable real-time policing and routing.
Provide Quantifiable Reliability: Offer a controlled, per-class error rate rather than just aggregate accuracy, as the cost of misclassification varies by application.
Handle Unknowns: Recognize and abstain from classifying applications never seen during training, rather than forcing them into known classes.
Remain Maintainable: Adapt to concept drift (changing application versions, traffic mixes) without expensive, disruptive full retraining.

Current literature addresses these issues in isolation (e.g., early-exit methods lack precision guarantees; robustness methods lack calibration; drift handling often requires full retraining). No existing system unifies early commitment, per-class certified precision, unknown detection, and lightweight drift maintenance at the scale of hundreds of application classes.

Methodology: CAPE-Net

The authors propose CAPE-Net (Conformal, Adaptive, Packet-Efficient Network), a unified framework built on a compact recurrent backbone (GRU or LSTM) with 18 early-exit checkpoints. The core innovation is treating the early-exit structure not merely as an efficiency device, but as a multi-purpose inference instrument.

1. Base Architecture

Input: Per-Packet-Information (PPI) sequences of up to 30 packets, encoded by four channels: inter-packet time (IPT), direction (DIR), payload size (SIZE), and TCP PSH flag (0 for QUIC).
Backbone: A two-layer recurrent network (hidden size 256) shared across all checkpoints.
Checkpoints: 18 exit heads located at packet indices $K \in \{1, \dots, 15, 20, 25, 30\}$ .
Training: Joint training with cross-entropy loss at each exit and self-distillation from the deepest exit ( $K=30$ ) to guide earlier heads.

2. Three Coupled Mechanisms

A. Per-Class Conformal Precision Calibration (Contribution 1)
Instead of a global confidence threshold, CAPE-Net uses a Learn-Then-Test (LTT) procedure with Holm correction to certify a per-class confidence threshold ( $\tau_c$ ).

Goal: Guarantee that the precision of committed flows for class $c$ is at least 0.95 with finite-sample, distribution-free validity.
Mechanism: Flows commit only if the checkpoint $K \ge K^*_c$ (the earliest reliable commit point for that class) AND the predicted probability $\hat{p} \ge \tau_c$ . Otherwise, the flow defers to a deeper exit.

B. Native Unknown-Application Detection (Contribution 2)

Mechanism: Uses a free-energy score ( $E(x) = -\log \sum \exp(\text{logits})$ ) read at a fixed checkpoint ( $K=10$ ).
Gate Logic: A "confidence-first" gate ensures that flows confidently classified by the adaptive early-exit rule are committed immediately. Only flows that fail to commit are passed to the energy gate. If $E(x) \ge \theta$ , the model abstains, emitting an "unknown" verdict.
Constraint: The threshold $\theta$ is set to maintain an in-distribution false-positive rate of $\le 0.05$ .

C. Drift-Aware, Safe Maintenance (Contribution 3)

Monitoring: Per-class Cumulative Sum (CUSUM) monitors track shifts in the realized exit depth ( $K_{used}$ ) and realized precision relative to the calibrated baseline.
Repair: When a class is flagged for drift, a bias-only update is performed. Only the bias terms of the flagged class's exit heads are updated; all other weights (including the backbone and other classes) remain frozen.
Safety: This ensures the update cannot degrade the performance of non-flagged classes.

Key Results

Evaluation on TLS (CESNET-TLS-Year22)

Evaluated on a 180-class corpus with a strict temporal split (Train: W1-W38, Cal: W39, Test: W40, Drift: W41-W50).

Precision Guarantee: On the held-out test week (W40), the 0.95 precision guarantee held for 98.3% of classes (GRU) and 97.8% (LSTM). The realized precision of committed flows reached ~0.99.
Efficiency: The typical flow is classified from 7–8 packets. The most confident classes commit on as few as 3 packets. This achieves near-full-sequence accuracy (within 0.04 macro-F1 of the 30-packet model) at roughly 25% of the latency.
Unknown Detection: The energy detector achieves an AUROC of 0.87 (GRU) at a 5% false-positive rate, correctly flagging ~43% of held-out unknown flows.
Drift Maintenance: Over a 10-week drift horizon, the bias-only update consistently recovers a small margin in macro-F1 (e.g., +0.005) at a cost of ~16–19 seconds per update, slowing degradation without full retraining.

Evaluation on QUIC (CESNET-QUICEXT-25)

Retrained from scratch on a year-long QUIC dataset (47 classes).

Transferability: The design transfers successfully, maintaining the precision guarantee and unknown detection capabilities.
Protocol Differences: QUIC flows commit later (mean exit ~15 packets) due to larger initial padding and longer flows.
Drift Behavior: While the CUSUM monitor detects drift effectively, the bias-only repair recovers almost nothing on QUIC ( $\Delta \approx 0$ ). The authors attribute this to the nature of QUIC drift being representational (feature-to-class mapping shifts) rather than prior-based, which bias updates cannot fix.

Significance and Claims

The paper claims that CAPE-Net demonstrates that early, calibrated, and maintainable classification at ISP scale is achievable within a single architecturally disciplined model.

Key takeaways emphasized by the authors:

Unified Instrument: The early-exit structure is a multi-purpose tool that simultaneously enables efficiency, certified precision, unknown detection, and drift sensing.
Operational Trade-offs: The system explicitly trades coverage (deferring ~25% of TLS flows) for a rigorous, finite-sample precision guarantee, which is critical for high-stakes network operations.
Protocol Awareness: While the architecture is protocol-agnostic, the latency budget and maintenance strategy must be recalibrated per protocol (e.g., QUIC requires later commitment and potentially heavier retraining than TLS).
Safety: The bias-only maintenance loop provides a "safe" mechanism to slow degradation between full retrainings, ensuring that maintenance actions never harm non-flagged classes.

The authors conclude that while the lightweight maintenance loop is highly effective for TLS, its limitations on QUIC suggest that a robust operational policy requires continuous monitoring to determine when cheap fixes are exhausted and full retraining is necessary.