acceptodds
Under review as a conference paper at ICLR 2027

FlowGuard: A Machine Learning System for DDoS Mitigation via LLM-Augmented In-Switch Classification

Abstract

Distributed denial-of-service (DDoS) attacks remain a severe threat to network availability. Programmable switches enable line-rate DDoS mitigation, but their limited computational and memory resources restrict in-switch classifiers to lightweight models, which are both less accurate on attack traffic that resembles legitimate flows and quickly outdated as attack patterns evolve. To address these challenges, we present FlowGuard, a closed-loop hybrid DDoS mitigation framework that combines line-rate in-switch classification with LLM-assisted out-of-switch analysis and classifier adaptation. In the switch data plane, a lightweight decision tree performs line-rate flow classification and enables immediate mitigation. Outside the switch, an LLM analyzes compact window-level traffic statistics to produce risk assessments, victim subnets, and rationales for flagged windows, from which FlowGuard derives flow-level weak labels. When statistical drift is detected, a second LLM stage determines whether retraining is warranted. If retraining is approved, FlowGuard trains a new decision tree from these weak labels together with the original labeled data, and deploys it only after it passes a safety gate. Through this closed-loop design, FlowGuard combines line-rate DDoS mitigation with the ability to adapt to evolving traffic patterns. Experiments on CIC-DDoS2019 and IoT-23 show that FlowGuard achieves an F1 score of 0.968, representing relative improvements of 24.6% and 43.8% over our static in-switch decision-tree and the in-network baseline SISTAR, respectively. Code and data are available at https://anonymous.4open.science/r/FlowGuard.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
DDoS detection, programmable switches, large language models, distribution shift, in-network machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vathOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FlowGuard combines a depth-4 in-switch decision tree with off-switch LLM analysis of traffic-window summaries and drift-triggered tree adaptation. The window stage detects suspicious traffic and identifies victim subnets, which are used to derive weak flow labels; a second LLM decides whether a drift event warrants retraining, and a gate compares candidate and deployed trees on held-out weak labels and original labeled data. Offline experiments on streams composed from CIC-DDoS2019 and IoT-23 report window-level F1 of 0.968 on the development evaluation stream and 0.894 on a frozen validation stream. Adaptation improves the deployed tree on the former stream, while separate measurements characterize victim localization, packet-drop rates, and LLM inference cost.

Critical strengths
  1. Separating immediate packet-path classification from slower aggregate analysis is a sensible systems design. The paper details switch-compatible features, tree compilation, weak-label construction, and gated updates, and provides an accessible implementation. This makes the proposed closed loop more concrete than simply applying an LLM to individual traffic records.

  2. The evaluation includes useful controls and candid limitations: a same-digest rule, a simple volume baseline, a frozen validation stream, chronological test-then-train replay, and retraining-policy/sensitivity studies. Reporting actual packet-drop fractions separately from window alarms is particularly valuable, as is documenting that frequent retraining can degrade performance even with largely correct weak labels.

Critical weaknesses
  1. The incremental value of the LLM stages is not yet convincing relative to simpler alternatives. The detection F1 is 0.968 versus 0.961 for the fixed volume threshold; an evaluation-tuned volume threshold achieves 0.995 there and 0.889 on validation, versus 0.894 for the LLM (Section 4.2). Packet-volume victim ranking already contains the target in 99.8% of explained windows and has higher exact-/24 accuracy than the LLM (Table 5). Retraining on every drift matches the LLM-triggered tree/pipeline scores on both streams, with the LLM saving one training attempt per stream (Table 8). These results show some efficiency and worst-segment benefits, but do not establish a substantial learning advantage over a well-controlled lightweight detector and adaptation policy on the same information. The large headline gains over static flow/packet classifiers conflate model choice with access to aggregate temporal information.

  2. The deployment evaluation omits key information and timing constraints. Accuracy uses all-packet offline digests, whereas the switch samples approximately 1/1024 of packets; the effect of sampling/sketch error on detection, weak labels, and adaptation is unmeasured. Full-digest detection takes 1.47 s per approximately 1.07 s window, and end-to-end queueing, update delay, and packet mitigation are not evaluated. The minimal digest is a useful alternative, but combined detection/explanation/adaptation serving still requires measurement under actual switch-exported traffic.

  3. The evaluated traffic is too narrow to establish robustness for the intended DDoS setting. Each attack segment has one source and one or two targets, leaving distributed sources, subnet-wide victims, and several digest components insufficiently tested. The frozen validation stream is valuable, but its 30.3% window false-alarm rate also raises questions about operational robustness across changing benign traffic. Broader independent attack captures and benign shifts are needed before the systems claims are convincing. Together with the limited demonstrated LLM advantage, these gaps motivate weak rejection.

Questions for authors
  1. Can the authors compare with a lightweight learned or adaptive window-level detector using the same digest, and isolate the benefit of LLM retraining decisions under matched weak-label and retraining budgets? Uncertainty across independent attack captures would help interpret the small validation gap.

  2. Can the authors report results using actual switch-exported sampled digests under distributed, high-churn traffic, including switch resources, export bandwidth, queueing latency, and time until effective packet mitigation? The repository README also states that every classified flow exports a flow_digest; please include this adaptation traffic in the overhead accounting.

  3. How does the safety gate protect against systematic victim-prefix labeling errors or benign changes outside the original labeled data? Since periodic updates still degrade performance despite the shared gate, an evaluation of accepted-update failures and a genuinely held-out labeled audit would clarify what safety the gate actually provides.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer vath [Part 1/2]

We thank the reviewer for the careful reading and the concrete requests. All changes in the revised PDF are highlighted in orange, and the new experiments are collected in Appendices H–J.

W1 / Q1. Learned detectors on the same digest

Using the same full digest, the LLM achieved higher F1 on both original streams than logistic regression and GBDT trained on CIC-DDoS2019 windows. We trained logistic regression and gradient-boosted trees (GBDT) on CIC-DDoS2019 digests and tested them on the shift to IoT-23. Because all three models receive the same window digest, this comparison separates the choice of model from access to aggregate temporal information. With the full digest on the original evaluation and validation streams, the LLM reached F1 of 0.968 and 0.894, GBDT reached 0.950 and 0.837, and logistic regression reached 0.555 and 0.655.

The gap widened on two further IoT-23 floods, a moderate-rate SYN flood in capture 34-1 and a high-rate UDP flood amid scanning in 43-1. With the same full digest for all methods, the LLM was the only detector to detect all four recorded attack episodes, reaching F1 of 0.927 and 0.865 (mean 0.896) against 0.301 for GBDT and at most 0.500 across all ten non-LLM baselines.

Full-digest logistic regression scored 0% recall on 43-1, while full-digest GBDT scored 0% recall on 34-1 and flagged every window on 43-1 (100% false-alarm rate). Seven of the ten non-LLM baselines reached at most 3.4% recall on 43-1, where heavy background scanning masks the volume surge, and the other three scored 0% recall on 34-1 (Appendix H, Table 13).

Manuscript changes. Sections 4.1, 4.3, and 5, and Appendix H.

W1 / Q1 (continued). Retraining under a shared budget

Under matched label and attempt budgets on the evaluation stream, the LLM trigger reached a tree F1 of 0.852 against 0.777 for retraining on every drift. We held the initial tree, weak-labeling rule, and safety gate fixed and varied only the trigger. With at most 668 admitted labels and three attempts, both triggers made two attempts. Retraining on every drift spent 398 labels at an early drift in 60-1 and had too few left for the later 35-1 shift. The LLM skipped the 60-1 drift and used 507 labels at 35-1. On validation, both triggers selected the same update under a cap of 793 admitted labels and two attempts.

With exactly 100 labels per attempt, both triggers used 200 labels on evaluation and reached the same tree and pipeline F1. The LLM trigger dropped fewer benign packets (0.648% against 0.736%). The two triggers also tied on validation (Appendix I, Table 16).

Manuscript changes. Section 4.2 and Appendix I.

W2 / Q2. Switch exports and end-to-end performance

Line-rate drops come from the compiled tree, which matches the offline tree on every evaluated flow. The compiled switch tree agrees with the offline tree on all training and evaluation-stream flows (Appendix B). Per-flow state uses a fixed table of 2162^{16} slots (Appendix B). The tree classifies flows in the data plane, while the window stage runs off-switch, so concurrent inference leaves the switch's export workload unchanged. The switch updates the packet and byte counters and the LogLog sketches on every IPv4 packet. Packet sampling therefore affects only the top-destination share, victim-prefix, and cross-server similarity fields of the full digest, and the minimal digest uses none of them.

On the evaluation stream, an average of 259 flows/s reached classification. Flow records and sampled packet records together need an estimated 0.0455 Mb/s of payload bandwidth on average and 1.49 Mb/s in the peak 1.074 s window (Appendix D).

On one worker, the minimal digest took 0.54 s per request and processed 1.85 windows/s against an arrival rate of 0.93, with validation F1 of 0.886 against 0.894 for the full digest. Detection scores in the paper use offline digests computed from all packets, and the timings cover LLM requests only. End-to-end measurements of queueing and update-to-mitigation delay are left to future work.

Manuscript changes. Section 4.2 and Appendix D.

AuthorsOpenReview ↗

Reply to Reviewer vath [Part 2/2]

W3. Traffic coverage and validation false alarms

On the validation stream, the full-digest LLM had a lower false-alarm rate than the fixed volume threshold, the threshold rule, and SISTAR. The full-digest LLM flagged 30.3% of the 211 DDoS-negative validation windows, compared with 39.3% for both the threshold rule and the volume threshold and 60.2% for SISTAR (Table 4). On four further DDoS-negative IoT-23 traces (16,918 windows), the original full-digest LLM, without recalibration, flagged 0.007% of windows on benign Somfy-01 and Somfy-04 traffic (4.02% for the fixed volume threshold), 0.09% on 3-1 (3.25%), and 7.3% under heavy scanning in 49-1, where the full-digest threshold rule flagged every window.

For attacks on many hosts, we built a multi-victim replay from capture 60-1. All tested logistic regression and GBDT variants scored 0% recall on this replay. The full-digest LLM flagged the first attack window and reached 9.6% recall overall. Independently captured distributed attacks are the next step for this part of the evaluation.

Manuscript changes. Section 4.4 and Appendix H.

Q3. What the safety gate protects

For each update, the gate limits the loss of balanced accuracy on the original labeled data to one percentage point relative to the deployed tree and requires improvement on held-out weak labels.

To inject systematic errors, we replaced each eligible victim prefix with the most frequent alternative /24 before generating both the fitting and gate labels. Over five seeds of LLM-triggered retraining on evaluation, enforcing the gate under these errors raised overall tree F1 from 0.451 to 0.547, with unchanged pipeline F1. With the original weak labels on validation, it raised pipeline F1 from 0.8885 to 0.8912.

We then scored every attempted update against ground-truth labels on flows from the next four adaptation intervals, excluding the whole retraining candidate pool. With corrupted labels, the gate accepted a candidate at one evaluation drift in three of five seeds even though balanced accuracy on the audited flows fell from 0.793 to 0.454. On validation, it accepted the update in all five seeds, and pipeline F1 stayed at 0.8885 with or without the gate (Appendix I).

Manuscript changes. Section 3.5 and Appendix I.

We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer vath,

Thank you for the careful and detailed review. In response, we added learned detectors on the same digest (Section 4.3, Table 13), compared retraining triggers under matched budgets (Section 4.2, Appendix I), accounted for the flow_digest export traffic (Appendix D), evaluated four additional attack-free traces (Appendix H), and tested the safety gate under injected labeling errors (Appendix I).

Please let us know if any point needs more detail. If these additions resolve your concerns, we would be grateful if you could re-evaluate the submission. Thank you very much.

Best regards, Authors

AuthorsOpenReview ↗

Brief note before the discussion closes

Dear Reviewer vath,

We appreciate the care you put into your review. Before the discussion closes today, we would welcome any feedback on our response and the additions in the revised PDF. We are happy to clarify anything that remains unclear.

Thank you very much.

Best regards, Authors

Reviewer_S29KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FlowGuard, a hybrid framework for DDoS mitigation that combines lightweight in-switch traffic classification with LLM-assisted traffic analysis and adaptive classifier updates. The proposed system consists of three main components. First, a decision-tree classifier is deployed in programmable switches to classify flows and perform immediate packet-level mitigation. Second, an external LLM analyzes window-level traffic digests to identify potential attacks and localize victim subnets. Third, a drift-triggered adaptation mechanism uses LLM-assisted decisions and weak labels to retrain the in-switch classifier, with a safety gate determining whether updated models should be deployed. The authors evaluate FlowGuard on traffic constructed from CIC-DDoS2019 and IoT-23. The LLM-based window detector achieves an F1 score of 0.968 on the evaluation stream, compared with 0.961 for a volume-threshold baseline. On the held-out validation stream, the method achieves an F1 of 0.894. The proposed adaptation mechanism improves the deployed decision tree from 0.777 to 0.852 F1 on the evaluation stream while reducing unnecessary retraining attempts.

Critical strengths

Strength 1

The paper addresses an important challenge in network security: maintaining efficient DDoS mitigation under evolving traffic distributions while respecting the computational constraints of programmable switches. The framework also considers practical deployment constraints, including switch memory limitations, model compilation, traffic summarization, and inference latency. These implementation details strengthen the practical relevance of the work.

Strength 2

The paper evaluates several components of the proposed system, including window-level detection, victim localization, retraining policies, and inference costs. An interesting observation is that the LLM-based window detector exhibits better worst-case performance across the evaluated attack segments than the deterministic baselines, despite relatively small differences in aggregate F1.

Critical weaknesses

Weakness 1

The central contribution of the paper is the integration of LLM-assisted traffic analysis and adaptation into an in-switch DDoS mitigation framework. However, the experiments do not convincingly demonstrate that an LLM is necessary to obtain the reported improvements. In Table 1, the LLM-based window detector achieves an F1 score of 0.968, compared with 0.961 for a simple volume-threshold baseline. Although the authors provide evidence of better worst-case performance across attack segments, the aggregate improvement is relatively small considering the additional inference costs and system complexity. 

Weakness 2

The paper motivates FlowGuard using large-scale, rapidly evolving distributed DDoS attacks. However, the evaluation scenarios are substantially more restricted than this motivating setting. Section 4.4 explicitly acknowledges that each attack segment originates from a single source address and targets only one or two hosts. Consequently, important components of the traffic digest, including source distribution, cross-server similarity, and victim-ranking features, are not fully exercised. Furthermore, Section 4.1 states that two components of the traffic digest were introduced after inspecting the evaluation stream. Although the authors subsequently evaluate a frozen configuration on a validation stream, this stream includes a later portion of an IoT-23 capture already used in the evaluation. Thus, the experimental evidence for generalization to genuinely unseen attack families and independently collected traffic remains limited. Additionally, the full-digest detection stage requires an average of 1.47 seconds per window, exceeding the approximately 1.07-second window arrival interval on the evaluated hardware. Although the authors provide a faster minimal-digest alternative, the complete system has not been validated under sustained real-time operation.

Questions for authors

Q1

Table 1 shows that the LLM-based window detector achieves an F1 of 0.968, compared with 0.961 for a simple volume-threshold baseline. Although the LLM demonstrates better worst-case performance across attack segments, the aggregate improvement is relatively small. Can you provide comparisons with stronger non-LLM alternatives, such as lightweight supervised classifiers operating on the same window-level features? 

Q2

Section 4.1 acknowledges that two digest components were introduced after inspecting the evaluation stream, while the held-out validation stream includes a later portion of an already evaluated IoT-23 capture. And, Section 4.4 notes that each attack segment originates from a single source address and targets only one or two hosts. Could the authors clarify how the approach is expected to generalize to genuinely unseen attack families, large-scale distributed attacks, and heterogeneous benign traffic? 

Q3

Section 4.1 states that detection accuracy is evaluated using traffic digests computed from all packets rather than sampled switch exports. Meanwhile, Section 4.2 reports that full-digest LLM inference takes 1.47 seconds per window, exceeding the window arrival interval of approximately 1.07 seconds on the evaluated hardware. Could the authors clarify how sampling errors, switch-export overhead, and concurrent LLM inference would affect the end-to-end system? Can the reported detection accuracy and latency be maintained when using actual switch-exported digests under realistic traffic rates? 

Q4

Table 5 reports that a simple "Most packets /24" heuristic identifies the exact victim subnet in 99.8% of evaluated windows, compared with 98.1% for the LLM-based explanation stage. Could the authors clarify whether the LLM's advantage emerges under more challenging multi-victim or distributed attack scenarios, which are not covered by the current evaluation?

6 comments
AuthorsOpenReview ↗

Reply to Reviewer S29K [Part 1/2]

We thank the reviewer for the detailed questions and for noting the worst-segment results. All changes in the revised PDF are highlighted in orange, and the new experiments are collected in Appendices H–J.

W1 / Q1. Detection compared with stronger baselines

Using the same full digest, the LLM achieved higher F1 on both original streams than logistic regression and GBDT trained on CIC-DDoS2019 windows. We trained logistic regression and gradient-boosted trees (GBDT) on CIC-DDoS2019 digests and tested them on the shift to IoT-23. With the full digest on the original evaluation and validation streams, the LLM reached F1 of 0.968 and 0.894, GBDT reached 0.950 and 0.837, and logistic regression reached 0.555 and 0.655.

The gap widened on two further IoT-23 floods, a moderate-rate SYN flood in capture 34-1 and a high-rate UDP flood amid scanning in 43-1. With the same full digest for all methods, the LLM was the only detector to detect all four recorded attack episodes, reaching F1 of 0.927 and 0.865 (mean 0.896) against 0.301 for GBDT and at most 0.500 across all ten non-LLM baselines.

Full-digest logistic regression scored 0% recall on 43-1, while full-digest GBDT scored 0% recall on 34-1 and flagged every window on 43-1 (100% false-alarm rate). Seven of the ten non-LLM baselines reached at most 3.4% recall on 43-1, where heavy background scanning masks the volume surge, and the other three scored 0% recall on 34-1 (Appendix H, Table 13).

Manuscript changes. Sections 4.1, 4.3, and 5, and Appendix H.

W2 / Q2. Evaluation scope and generalization

The frozen original detector was evaluated on a later non-overlapping attack episode and on four additional DDoS-negative traces. For unseen traffic, the validation stream combines CIC-DDoS2019 capture files disjoint from the evaluation stream with a later, non-overlapping episode of capture 35-1 whose floods target two other hosts. On four further attack-free IoT-23 traces (16,918 windows), the original full-digest LLM, without recalibration, flagged 0.007% of windows on benign Somfy-01 and Somfy-04 traffic (4.02% for the fixed volume threshold), 0.09% on 3-1 (3.25%), and 7.3% under heavy scanning in 49-1, where the full-digest threshold rule flagged every window.

For attacks on many hosts amid competing background traffic, we built a multi-victim replay from capture 60-1 alongside benign bulk transfers and scanning. Its traffic rises in two steps, each below the 10× surge-hold threshold, so the rolling baseline absorbs the flood. All tested logistic regression and GBDT variants scored 0% recall on this replay. The full-digest LLM flagged the first attack window and reached 9.6% recall over the episode. Independently captured distributed attacks are the next step for this part of the evaluation.

On the original evaluation stream, drift-triggered retraining raised tree F1 on the 35-1 segment from 0.000 to 0.461.

Manuscript changes. Section 4.4 and Appendix H.

W2 / Q3. Sampling and latency in deployment

Packet drops in the switch do not wait for sampled exports or LLM inference. The compiled switch tree agrees with the offline tree on all training and evaluation-stream flows (Appendix B). The tree classifies flows in the data plane, while the window stage runs off-switch, so concurrent inference leaves the switch's export workload unchanged. The switch updates the packet and byte counters and the LogLog sketches on every IPv4 packet. Packet sampling therefore affects only the top-destination share, victim-prefix, and cross-server similarity fields of the full digest, and the minimal digest uses none of them.

On the evaluation stream, the estimated payload bandwidth for sampled packet records and per-flow adaptation records combined was 0.0455 Mb/s on average and 1.49 Mb/s in the peak 1.074 s window (Appendix D).

On one worker, the minimal digest took 0.54 s per request and processed 1.85 windows/s against an arrival rate of 0.93, with validation F1 of 0.886 against 0.894 for the full digest. Detection scores in the paper use offline digests computed from all packets, and the timings cover LLM requests only.

Manuscript changes. Section 4.2 and Appendix D.

AuthorsOpenReview ↗

Reply to Reviewer S29K [Part 2/2]

Q4. Victim localization in more challenging traffic

On the constructed replays, we evaluated victim-ranking rules on known attack windows, and the rule that recovers the victim differs between the two replays. On the original evaluation stream, packet-count ranking identified the exact victim /24 in 99.8% of explained attack windows (Table 5). In the multi-victim UDP replay, the four prefixes with the most packets recovered 99.88% of victims on average. In the SYN flood replay with a large benign HTTPS transfer, packet count alone ranked the victim first in 19.05% of attack windows, and weighting packet count by log(1 + flows) raised this to 95.24%. The LLM explanation stage returns a victim subnet together with an incident explanation.

Manuscript changes. Appendix H, "Victim ranking on known attack windows."

We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer S29K,

Thank you again for your review and for noting the worst-segment results. The main additions in the revision are as follows:

  • Non-LLM baselines on the same features: Section 4.3 and Table 13.
  • Additional attack-free traces: Appendix H.
  • Export overhead and inference throughput: Section 4.2 and Appendix D.

We remain available for any further questions during the discussion. If our response has addressed your concerns, we would be grateful if you would consider raising your score. Thank you very much.

Best regards, Authors

Reviewer_88hhOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a pipeline for DDoS attack detection that combines a lightweight decision tree running directly in a programmable switch with an LLM analyzing aggregate per-window traffic statistics exported by the switch. The LLM flags attack windows and identifies the victim subnet, from which weak labels are derived; when a statistical drift detector fires, a second LLM call decides whether the tree should be retrained on these weak labels to follow evolving traffic patterns, with a safety gate before deployment. The evaluation uses a stream shifting from CIC-DDoS2019 to IoT-23.

Critical strengths

The paper tackles a relevant problem: adapting resource-constrained in-switch classifiers to evolving DDoS traffic is practically important, and the paper addresses it end to end. In addition, I appreciated the detailed and honest empirical reporting: the evaluation includes simple baselines, ablations, and inference costs, and openly discloses design choices made on the evaluation stream and the limits of the results.

Critical weaknesses

My main concern it that the proposed pipeline combines established components, including in-switch trees, LogLog digests, Wasserstein-based drift detection, heuristic weak labels, and a prompted LLM. Furthermore, the closed loop of in-network inference and foundation-model-assisted retraining was already proposed in Caravan (OSDI '24). Overall, I belive that, especially from a technical perspective the novelty is limited.

Another concern is that the LLM operates on a low-dimensional vector of aggregate statistics with well-defined semantics, for which a conventional classifier would be a natural baseline. Moreover, the paper's own results suggest that simpler methods may suffice: a one-parameter volume threshold achieves an F1 score of 0.961 versus 0.968 for the proposed approach, a digest using only cardinality and volume improves the F1 score to 0.992, prefix ranking by packet volume localizes the victim in 99.8% of windows, and LLM-triggered retraining performs comparably to retraining at every drift, so the need of introducing an LLM is not clear.

Questions for authors

Here some questions:

  1. How does a conventional classifier (e.g., gradient-boosted trees) trained on the CIC-DDoS2019 digests perform on the IoT-23 shift, compared to an LLM?
  2. How does FlowGuard differ from Caravan, which also retrains in-network models using foundation-model labels? Could Caravan be included as a baseline?
  3. Given that the cardinality and volume-only digest achieves the best F1 (0.992) and meets the real-time budget, which of the remaining digest features are needed, and on what traffic?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer 88hh [Part 1/2]

We thank the reviewer for the specific questions. All changes in the revised PDF are highlighted in orange, and the new experiments are collected in Appendices H–J.

W1 / Q2. Pipeline contribution and comparison with Caravan

Caravan relies on its knowledge sources to label each sampled flow correctly, whereas FlowGuard derives flow labels from a window-level verdict and victim subnet that the LLM infers from statistics across flows (Appendix J, Table 17). Under distribution shift, benign and attack flows from the same traffic can look alike on per-flow features. The Caravan paper itself reports that inaccurate knowledge sources can turn the accuracy gain from retraining negative (its Section 5.2.3).

We added a Caravan-style batched flow labeler and a Caravan-inspired ungated update policy as baselines. For labeling, we adapted Caravan's batched flow-labeling prompts to our flow features and used the same Qwen checkpoint. On flows labeled by both methods, FlowGuard reached weak-label F1 of 0.987 against 0.616 on evaluation and 0.971 against 0.286 on validation (Table 18). Both labelers found every attack flow, but the Caravan-style labeler also flagged nearly every benign flow in capture 35-1 as an attack. With the tree learner and safety gate fixed and all available labels from each method used, FlowGuard's labels raised overall evaluation tree F1 to 0.852, against 0.777 with Caravan-style labels, and 35-1 tree F1 to 0.461, against 0.000. This held under both LLM-triggered and on-drift updates.

Caravan uses its LLM only for labeling, and detection stays with the data-plane classifier. In FlowGuard, the LLM also runs alongside the switch tree as a live window-level detector that names the victim subnet and provides incident explanations. To measure what the window detector adds, we held FlowGuard's switch tree fixed. Adding the LLM window detector to FlowGuard's static switch tree raised detection F1 from 0.777 to 0.968 on evaluation and from 0.437 to 0.893 on validation (Tables 1 and 4).

The two systems also differ in how labels are produced and how updates are deployed. Caravan labels flows by batching flow vectors with 10 to 20 labeled examples or by reusing a rule function the LLM writes from example flows, retrains on a class-balanced buffer, and deploys the result directly. FlowGuard keeps the original labeled data at full weight, down-weights the weak labels, and deploys a candidate only if it improves balanced accuracy on held-out weak labels while staying within one percentage point of the deployed tree on the original data.

For updates, we replayed a Caravan-inspired policy that retrains on a class-balanced buffer and deploys each tree without a gate. With on-drift updates on evaluation, the ungated policy more than tripled benign packet drops compared with gated on-drift FlowGuard (1.92% against 0.60%) and lowered pipeline F1 from 0.968 to 0.960. On validation, the ungated on-drift tree reached an F1 of 0.428 against 0.442 for gated on-drift FlowGuard and dropped 35.9% of benign packets. On both streams, gated on-drift FlowGuard achieved higher pipeline F1 and fewer benign drops than either ungated schedule (Table 19).

Manuscript changes. Sections 2, 4.1, and 4.2, and Appendix J.

W2 / Q1. Learned detectors on the same digest

Using the same full digest, the LLM achieved higher F1 on both original streams than logistic regression and GBDT trained on CIC-DDoS2019 windows. We trained logistic regression and gradient-boosted trees (GBDT) on CIC-DDoS2019 digests and tested them on the shift to IoT-23. With the full digest on the original evaluation and validation streams, the LLM reached F1 of 0.968 and 0.894, GBDT reached 0.950 and 0.837, and logistic regression reached 0.555 and 0.655.

The gap widened on two further IoT-23 floods, a moderate-rate SYN flood in capture 34-1 and a high-rate UDP flood amid scanning in 43-1. With the same full digest for all methods, the LLM was the only detector to detect all four recorded attack episodes, reaching F1 of 0.927 and 0.865 (mean 0.896) against 0.301 for GBDT and at most 0.500 across all ten non-LLM baselines.

Full-digest logistic regression scored 0% recall on 43-1, while full-digest GBDT scored 0% recall on 34-1 and flagged every window on 43-1 (100% false-alarm rate). Seven of the ten non-LLM baselines reached at most 3.4% recall on 43-1, where heavy background scanning masks the volume surge, and the other three scored 0% recall on 34-1 (Appendix H, Table 13).

Manuscript changes. Sections 4.1, 4.3, and 5, and Appendix H.

AuthorsOpenReview ↗

Reply to Reviewer 88hh [Part 2/2]

W2 (continued). Retraining trigger

Under a shared budget of labels and attempts, the LLM trigger reached a tree F1 of 0.852 against 0.777 for retraining on every drift. We held the initial tree, weak-labeling rule, and safety gate fixed and varied only the trigger. With at most 668 admitted labels and three attempts, both triggers made two attempts. Retraining on every drift spent 398 labels at an early drift in 60-1 and had too few left for the later 35-1 shift. The LLM skipped the 60-1 drift and used 507 labels at 35-1. On validation, both triggers selected the same update under a cap of 793 admitted labels and two attempts.

Manuscript changes. Section 4.2 and Appendix I.

Q3. When the richer digest helps

Adding victim-prefix features to the minimal digest improves detection on the recorded UDP flood with scanning and reduces false alarms on the benign bulk-download trace. For isolated floods, cardinality and volume alone reached F1 of 0.992 on the original evaluation stream and 1.000 on the additional SYN flood. On the five additional traces, we added each feature group to the minimal digest and removed it from the full digest, with the prompt and scored windows fixed.

On the UDP flood amid scanning, adding victim-prefix features raised F1 from 0.387 to 0.885 and lowered the false-alarm rate from 7.2% to zero. On a benign bulk download, they cut false alarms from 53.7% to 6.1%. Flow-structure features also helped on the UDP flood, raising F1 to 0.767 (Table 15).

Manuscript changes. Section 4.3 and Appendix H.

We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 88hh,

Thank you for your time on our submission. On the relation to Caravan, Section 2 now explains the difference, and Appendix J compares the label sources (Table 18). On the role of the LLM, Section 4.3 adds learned detectors on the same digest (Table 13) and shows on which traffic the richer digest features improve detection, and Section 4.2 compares retraining triggers under a shared budget.

We would be happy to clarify anything further before the discussion period ends. If these revisions address your concerns, we would appreciate it if you could re-evaluate the paper. Thank you very much.

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and discussion

Dear Area Chair and Reviewers,

Thank you for your time and the constructive reviews. To help your assessment, we summarize how the revision addresses the main concerns.

Confirmed by Reviewer 88hh as answering the original questions

  1. Learned detectors on the same digest (88hh, S29K, vath). The LLM outperforms GBDT and logistic regression on both original streams (0.968 and 0.894 F1, against 0.950 and 0.837 for GBDT) and is the only detector that detects all four additional attack episodes (Section 4.3, Table 13).
  2. Comparison with Caravan (88hh). Caravan-style per-flow labeling marks 94% of benign flows as attacks, against 2% for FlowGuard's window-level labels, which let the switch tree recover the missed 35-1 segment (0.000 to 0.461 tree F1, Appendix J).

Addressed with new results

  1. Retraining under a shared budget (88hh, vath). On the evaluation stream, LLM-triggered retraining reaches 0.852 tree F1 against 0.777 for retraining on every drift (Section 4.2, Appendix I).
  2. Benign traffic outside the tuning data (S29K, vath). On four attack-free IoT-23 traces, the frozen detector flags 1 of 14,704 windows on benign device traffic and 7.3% under heavy scanning, where the full-digest threshold rule flags every window (Appendix H).
  3. Export overhead and latency (S29K, vath). Export records are estimated at 0.0455 Mb/s on average, and the minimal digest keeps pace with the window rate on one worker (Appendix D).
  4. Safety gate (vath). Under injected labeling errors on the evaluation stream, the gate raises tree F1 from 0.451 to 0.547 (Appendix I).

Final comments

In reply to Reviewer S29K's final comment, we added that every learned baseline scores 0% recall on the multi-victim replay while FlowGuard's recall does not fall as more victims are attacked, and that the LLM names a subnet containing the target in every explained attack window (Table 5).

Thank you very much.

Best regards, Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.