AdvNet: Searching for Adversarial Traces

AdvNet searches for adversarial traces using evolutionary algorithms. The main entry point for running the search is:

search_adv_traces.py

This script initializes the search process and runs the optimization algorithm to find traces that maximize the target score defined in a user-provided evaluation function.

---

Running AdvNet

To start searching for adversarial traces, run:

python search_adv_traces.py [arguments]

Arguments

--type

A unique integer identifier representing the domain in which adversarial traces are being searched.

Inside search_adv_traces.py, each domain corresponds to a block:

if args.type == X:

If you want to use AdvNet for a new domain, you must add a new if block with a unique type.

---

--alg

Specifies the algorithm used for searching adversarial traces.

Currently supported:

0 -> Random Generation (RG)
1 -> Genetic Algorithm (GA)

---

--trace_length

Length of the trace to search.

Example:

--trace_length=5

---

--l_bounds

Lower bound for each dimension of the trace.

Example:

--l_bounds 1 1 1 1 1

---

--u_bounds

Upper bound for each dimension of the trace.

Example:

--u_bounds 10 10 10 10 10

---

--seed

Random seed for reproducibility.

Example:

--seed=42

---

--total_time

Total search time in seconds.

Example:

--total_time=3600

---

--initial_pop_file

Optional file containing an initial population of traces to start the search.

--initial_pop_file=file_name

---

Evaluation Function

Users must implement an evaluation function that computes the score of a given trace.

The function will be called by AdvNet during the search.

Required Input

The evaluation function must accept the trace. Currently we support only 1D array and integer values.

---

Additional Parameters

The evaluation function can accept additional parameters beyond the trace.

To pass those parameters, include them when instantiating the CCProblem class.

Example location:

search_adv_traces.py
    elif args.type == 1:
        elif args.alg == 1:
            if args.initial_pop_file == "None":

When calling CCProblem, pass the additional arguments at the end.

Example:

problem = CCProblem(..., extra_param1, extra_param2)

These parameters will be forwarded to the evaluation function via *args.

---

Adding a New Domain

To use AdvNet in a new domain, follow these steps.

---

Step 1: Add a New Domain Block

Open:

search_adv_traces.py

Add a new block:

if args.type == NEW_TYPE:

Inside this block:

• Configure the algorithm
• Instantiate CCProblem
• Provide required parameters

---

Step 2: Modify GA/problem.py

Open:

GA/problem.py

Inside the _evaluate method, add a block for your domain:

if self.type == NEW_TYPE:

Call your evaluation function here.

Example:

score = evaluate(x, *args)

Where:

• x is the trace (NumPy 1D array)
• *args are the additional parameters passed from CCProblem

---

Step 3: Logging

Inside GA/problem.py, implement logging for the search process.

update_max_score

This method logs only when a better trace is discovered.

update_max_score(...)

Use it to record improvements in the best score found so far.

---

log

This method logs every trace evaluated during the search.

log(...)

This provides a full history of the search process.

---

Example Command

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

---

Summary

To extend AdvNet to a new domain:

1. Add a new args.type block in search_adv_traces.py
2. Implement your evaluation function
3. Pass any additional parameters via CCProblem
4. Add a new case in GA/problem.py::_evaluate
5. Implement logging via update_max_score and log

Using Modified Mahimahi

We modified Mahimahi to support time-varying latency and additional logging. The repo is available here:
Modified Mahimahi (delay-trace branch)