DoHFuse: A Dual-Branch Architecture with DMAGLSTM for Website Fingerprinting over DNS over HTTPS/3

arXiv:2606.24105 · cs.CR · Submitted 2026-06-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: I'm Nadia.

Elias: And I'm Elias, here with Priya.

Nadia: We're looking at the paper: "DoHFuse: A Dual-Branch Architecture with DMAGLSTM for Website Fingerprinting over DNS over HTTPS/3".

Elias: Here's the gist of the paper. DoHFuse proposes a dual-branch learning framework with DMAG-LSTM to investigate whether DNS over HTTPS/3 traffic remains susceptible to Website Fingerprinting attacks even with padding and multiplexing.

Nadia: Problem Statement Website fingerprinting (WF) attacks pose a severe threat to online privacy by allowing passive adversaries to infer user activity through encrypted traffic patterns.

Elias: This threat is particularly concerning in Internet of Things (IoT) environments where devices rely on encrypted communications but often lack strong endpoint protections.

Nadia: The study addresses the critical gap regarding DoH/3 protocol vulnerability, as there is no systematic investigation of such attacks under this protocol.

Paper summary: Elias: So wrapping up this paper "DoHFuse: A Dual-Branch Architecture with DMAGLSTM for Website Fingerprinting over DNS over HTTPS/three", the authors are essentially saying that we need to pay close attention to the timing structure of DoH over HTTP three <ref:2606.24105#pg1,DoHFuse: A Dual-Branch Architecture with DMAGLSTM for Website Fingerprinting over DNS>.

Nadia: They’ve shown that their dual-branch model, DoHFuse, can do a better job identifying website fingerprints than existing methods when dealing with this specific protocol.

Priya: What this means for us is that simply applying padding strategies isn't enough to secure privacy in these encrypted DNS environments. The timing is the key part here.

Elias: I think the title itself, DoHFuse, hints at how they combine those two different modeling approaches—the temporal sequence and the statistical features—to get a more robust result.

Nadia: And the fact that they built this dataset from real network environments gives it a lot of practical weight when we think about defending IoT devices.

Priya: It moves the conversation toward building defenses that are specifically aware of these burst-gap request patterns, rather than just looking at simple packet lengths.

Elias: That’s the main implication: if you’re working on DoH over HTTP three security, you have to focus on timing leakage as a primary defense target.

Conclusion: Nadia: So, we’re wrapping up this look at DoHFuse, which is about using a dual-branch model with DMAGLSTM to see if DNS over HTTPS three traffic still leaks website fingerprints even with padding and multiplexing.

Elias: Yeah, so the core idea here is testing that specific protocol under these conditions. The authors put together this framework to investigate whether those common countermeasures actually work against this kind of tracking.

Priya: What’s interesting is how they handle the data input; they use a sequence of inter-arrival times and some coarse statistical bits, which sounds like they're looking at timing patterns versus just looking at the packet size.

Nadia: Right, and their result shows an accuracy of eighty-eight point zero five percent in this closed world test with four hundred forty nine classes. That’s a pretty solid number for a privacy attack model.

Elias: It confirms what we suspected about the protocol itself; DoH over HTTP three traffic remains susceptible to website fingerprinting attacks, so padding alone isn't enough protection.

Priya: But what this really means for us is that if you're building defenses, you can’t just rely on padding mechanisms anymore because the timing is still a decisive signal here.

Nadia: Exactly; the paper suggests that the most important thing in DoH over HTTP three traffic is actually burst-gap request dynamics, which means timing leakage is what we need to target for defense.

Elias: That points toward a necessary change in how we think about anonymity in these environments; it shouldn't be assumed by default, but defenses need to explicitly focus on that timing leak.

Priya: So the implication is that privacy protection needs to be more surgical, targeting those specific temporal patterns rather than just generic encryption.

Nadia: It’s a pretty strong finding because they built this benchmark dataset from real-world traces from four hundred forty nine websites, which gives it some real practical weight.

Elias: And the limitation they mentioned is that packet size by itself isn't reliable anymore because of all the padding and encryption in this setting.

Priya: That makes sense, because if you can’t trust one signal, you have to build a model that understands the other signals too, which is what their dual-branch approach does.

Nadia: We’ve seen how they built it and what the numbers showed; this pushes us to look at browser-agnostic modeling next.

ZunDong Zhang, Yanan Cheng, Zhaoxin Zhang, Xueyang Huo, Changjiang Wu

Harbin Institute of Technology · China Mobile Research Institute

cs.CR

Submitted: 2026-06-23

Updated: 2026-06-23

DOI: 10.1109/JIOT.2026.3734028

Code: https://github.com/grasstractor/DoHFuse

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 82/100

The gist: DoHFuse proposes a dual-branch learning framework with DMAG-LSTM to investigate whether DNS over HTTPS/3 traffic remains susceptible to Website Fingerprinting attacks even with padding and…

Key concepts

Website Fingerprinting (WF)
This attack lets passive observers guess a user's activity by analyzing patterns in encrypted traffic. It works by looking at subtle differences in how a website communicates, even when the data is scrambled. This poses a privacy risk because it can reveal browsing habits without needing to decrypt the traffic.
DoH/3 Protocol
This refers to using DNS over HTTPS with version 3 for name resolution. The study focuses on this specific encrypted communication method, examining if its structure allows attackers to infer user behavior through traffic patterns, even when techniques like padding are used.
DMAG-LSTM
This is an improved type of Long Short-Term Memory network used in the model's temporal branch. It is designed specifically to capture complex, burst-aligned temporal patterns within the sequence of request intervals (Δt), helping the system detect timing leakage that might otherwise be hidden.
Dual-Branch Learning Framework
This is a modeling approach where two separate pathways process data simultaneously. One branch looks at time sequences (temporal branch) using DMAG-LSTM, while the other analyzes global statistics (statistical branch). The final features are then combined to make a more robust prediction about website identity.

Terminology

Summary

DoHFuse proposes a dual-branch learning framework with DMAG-LSTM to investigate whether DNS over HTTPS/3 traffic remains susceptible to Website Fingerprinting attacks even with padding and multiplexing. p. 5

Problem Statement

Website fingerprinting (WF) attacks pose a severe threat to online privacy by allowing passive adversaries to infer user activity through encrypted traffic patterns. This threat is particularly concerning in Internet of Things (IoT) environments where devices rely on encrypted communications but often lack strong endpoint protections. p. 0 The study addresses the critical gap regarding DoH/3 protocol vulnerability, as there is no systematic investigation of such attacks under this protocol. p. 0

Data and Methodology

The research involves collecting and publicly releasing the first real-world benchmark dataset of DoH/3 traffic, comprising traces from 449 websites. The proposed DoHFuse architecture is a dual-branch model that integrates inter-arrival time sequences and refined statistical features through an improved DMAG-LSTM. p. 0 The input to the model consists of a sequence of inter-request intervals (Δt) and a 2-D vector of coarse statistical attributes. p. 4

Model Architecture Details

DoHFuse is organized into five sequential components: (1) Input (Packets), (2) Feature Engineering, (3) Temporal Branch, (4) Statistical Branch, and (5) Feature Fusion. The temporal branch utilizes DMAG-LSTM to capture burst-aligned temporal patterns. p. 0 Simultaneously, the statistical branch processes global indicators like the total number of client-to-resolver request packets and the coefficient of variation (CV). p. 4

Key Findings and Contributions

Experimental results demonstrate that DoHFuse achieves an accuracy of 88.05% in a closed-world setting of 449 classes. p. 0 The study confirms that DoH/3 traffic remains susceptible to WF attacks, suggesting that commonly deployed padding mechanisms alone are insufficient for privacy protection. p. 0 The model's performance is superior to state-of-the-art baselines under a unified protocol.

Limitations and Future Directions

A limitation noted is that packet size alone is no longer a reliable signal in the DoH/3 setting due to padding and encryption. Furthermore, analysis of sparse websites reveals that sites with minimal DNS interactions generate patterns too indistinct for accurate identification. Future work is suggested to explore browser-agnostic modeling and defense-aware evaluation against active countermeasures.

Conclusion

The work establishes that the decisive signal in DoH/3 lies in timing specifically, namely the burst-gap request dynamics. The practical implication is that DoH/3 should not be assumed to deliver anonymity by default; effective defenses must explicitly target timing leakage.

Acknowledgments

This work is supported by the Shandong Provincial Natural Science Foundation(Grant No. ZR2024QF138), the S&T Program of Shijiazhuang (Grant No. XXXX), and the Scientific Research Innovation Fund of Harbin Institute of Technology. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Increase accuracy in website fingerprinting detection by designing a dual-branch architecture that fuses inter-arrival time sequences with refined statistical features, as demonstrated by DoHFuse achieving an accuracy of 88.05% in a closed-world setting of 449 classes.

  2. Develop robust open-world detection capabilities by implementing a probability-thresholding strategy based on the maximum softmax probability (MSP), allowing the model to reject unknown classes without requiring any additional training.

  3. Enhance temporal modeling of burst and gap structures in DoH/3 traffic by utilizing the DMAGLSTM module, which employs multi-scale forget-gate fusion to capture both microbursts and long-gap timing patterns.

  4. Improve model generalization across different class granularities by employing a late soft fusion strategy in the feature fusion stage, allowing each branch to preserve its inductive bias (local timing vs. global traffic shape) before integration.

  5. Create a standardized and reproducible research environment by publicly releasing the first dedicated DoH/3 WF dataset and providing the raw packet captures (PCAP) and the derived per-visit CSV statistics for direct reuse.

Sources

Related papers