AdvNet

Revealing performance issues in network protocols by generating adversarial environments.

Overview

Congestion-control (CC) algorithms are expected to deliver good performance across the enormous diversity of real Internet conditions — but testing them by hand only covers the scenarios a human thinks to try. AdvNet automatically generates the network environments that make a protocol look bad. It frames "find a hard environment" as an optimization problem and uses machine-learning-based search to evolve time-varying link traces (bandwidth, latency, queue size) that maximize the performance gap between a target protocol and a reference, with a built-in noise-handling mechanism to keep the search reliable on a noisy emulator.

As shown in the figure above, AdvNet runs a closed loop: an Adversarial Trace Generator proposes a trace, an Execution Module emulates it for both the reference and the target, a Scoring component turns the result into an adversarial score, and a Post-Learning Selection stage picks the strongest discovered traces. Applied to the Linux kernel, AdvNet surfaced previously unnoticed CC bugs and hidden limitations across many implementations.

Key findings (§4.3 — Comparison of TCP Protocols)

AdvNet was used to run pairwise robustness comparisons of all 17 TCP congestion-control algorithms in the Linux kernel (2-hour search budget per pair; the final 10% of the time reserved for the post-learning selection phase):

• Every protocol is vulnerable — each one suffers a large performance drop under some adversarial environment that AdvNet discovers.
• Vulnerability discovery is reference-dependent — which environments expose a weakness depends strongly on the reference protocol (e.g. with bbr as reference at tcoeff=0.5, 6 of 16 protocols produce a large positive score).
• cdg is the most robust when balancing high throughput against low latency; lp and htcp are preferable when latency is ignored.
• The ranking is highly sensitive to the throughput-vs-latency weighting tcoeff. At tcoeff=1 (zero weight on latency) a specific bbr bug (§4.4.1) pushes every bbr-as-target scenario above 0.8, and delay-based vegas scores above 0.9.

Mean adversarial score of each protocol as the target under tcoeff=0.5 (x-axis) vs tcoeff=1 (y-axis). Lower is more robust.

See the paper for the full pairwise heatmaps and analysis.

Repository structure

Path	Description
search_adv_traces.py	Main entry point; dispatches a domain (--type) and a search algorithm (--alg).
GA/	Genetic-algorithm search (ga.py, problem.py, mutation.py).
random_generator/	Uniform random-search baseline (RG).
BO/, BL/	Bayesian-optimization and epsilon-greedy baseline searchers.
RL/	Reinforcement-learning agents and environments.
scoring/	Adversarial scoring of reference-vs-target performance.
single_cc/, mptcp/, picoquic/, dchannel/, multiflow/	Emulation backends for single-flow CC, MPTCP, QUIC, data channels, and multi-flow setups.
utils/, selection_simulator/	Helpers and the post-learning selection simulator.
Plots/	Figures/plotting scripts.

Installation

pip install -r requirements.txt

AdvNet drives a network emulator and (for the ns-3 backend) an ns-3 build:

• ns-3 backend: point AdvNet at your ns-3 source tree via an environment variable:

  export ADVNET_NS3_PATH=/path/to/ns-3-dev

• Local configuration (config.py): copy the template, then set parent_folder:

  cp config.example.py config.py

  config.py is gitignored and imported throughout the emulation backends. Its key setting is parent_folder — the absolute path (with a trailing slash) of the directory that contains the AdvNet/ repo as well as the emulator's packet-logs/ and packet-logs-2/ output directories. Backends read traces from <parent_folder>AdvNet/traces/ and read/write per-run logs under <parent_folder>packet-logs[-2]/. For example, if the repo lives at /home/you/AdvNet, set parent_folder = "/home/you/" and create the log dirs once:

  mkdir -p /home/you/packet-logs /home/you/packet-logs-2

• Modified Mahimahi: the emulation backends use a modified Mahimahi that supports time-varying latency and extra logging: Modified Mahimahi (delay-trace branch).

Usage

python search_adv_traces.py [arguments]

Argument	Description
--type	Integer id of the evaluation domain (each args.type block in search_adv_traces.py).
--alg	Search algorithm: 0 = Random Generation (RG), 1 = Genetic Algorithm (GA).
--trace_length	Number of dimensions in the trace to search.
--l_bounds / --u_bounds	Per-dimension lower/upper bounds (space-separated).
--seed	Random seed for reproducibility.
--total_time	Total search time, in seconds.
--initial_pop_file	Optional file with an initial population of traces.

Example

python search_adv_traces.py \
    --type=1 \
    --alg=1 \
    --trace_length=5 \
    --l_bounds 1 1 1 1 1 \
    --u_bounds 10 10 10 10 10 \
    --seed=42 \
    --total_time=3600

Adding a new domain

AdvNet is extensible to new protocols/environments. To add a domain:

1. Add a new if args.type == NEW_TYPE: block in search_adv_traces.py that configures the algorithm and instantiates CCProblem with your parameters.
2. Implement an evaluation function that takes a trace (a 1-D array of integers) plus any extra parameters and returns a scalar score. Extra parameters are passed positionally when constructing CCProblem and forwarded to your function via *args.
3. Add a matching if self.type == NEW_TYPE: case in GA/problem.py::_evaluate that calls your evaluation function.
4. Use update_max_score(...) to record improvements to the best trace, and log(...) to record every evaluated trace.

Citation

If you use AdvNet, please cite the published version:

@article{ahmed2026advnet,
  title     = {AdvNet: Revealing Performance Issues in Network Protocols by Generating Adversarial Environments},
  author    = {Ahmed, Shehab Sarar and Sentosa, William and Zhang, Yinjie and Lebendiker, Yoav and Shnaiderman, Mickey and Gilad, Tomer and Jay, Nathan H. and Godfrey, Brighten and Schapira, Michael},
  journal   = {Proceedings of the ACM on Networking (PACMNET)},
  volume    = {4},
  number    = {CoNEXT2},
  pages     = {1--22},
  year      = {2026},
  publisher = {Association for Computing Machinery},
  doi       = {10.1145/3808660}
}

Published in Proceedings of the ACM on Networking (CoNEXT 2026) · DOI · arXiv preprint.

License

Released under the MIT License.