Skip to content
KitploitKITPLOIT
ToolsExploitsBlog
Log in
Submit
ToolsExploitsBlog
Submit

Hacking, PenTest, and Cybersecurity Tools for Your Security Arsenal!

Kitploit is a directory of hacking, cybersecurity, and pentesting tools. Discover the latest project updates to find vulnerabilities, analyze systems, automate testing, and strengthen your security.

FeedsContactPrivacy© 2026 Kitploit

Tool Directory

Categories

View all categories
Loading categories
AGAS — LLM-driven agentic group shilling attack framework that manipulates black-box collaborative-filtering recommender rankings using adaptive multi-role strategies while evading detection. | Kitploit
Tools/GitHubGitHub/phkhanhtrinh23/agas
Machine LearningPapers & ResearchLearning & EducationAI SecurityAdversarial Attack
GitHubphkhanhtrinh23/agas

AGAS

LLM-driven agentic group shilling attack framework that manipulates black-box collaborative-filtering recommender rankings using adaptive multi-role strategies while evading detection.

View Repository
21041 month agoNot yet reviewed
Share

An Efficient and Effective Agentic Group Shilling Attack on Recommender Systems

arXiv visitors

This is the official code to the paper: "An Efficient and Effective Agentic Group Shilling Attack on Recommender Systems". This paper introduces AGAS, which is an LLM-driven shilling attack against black-box collaborative-filtering recommenders. One Coordinator orchestrates a pool of fake-user workers over a sequence of rounds. In each round, the Coordinator picks one of eight strategies and assigns a role to every worker. Workers then decide which items to rate using their own ReAct-style reasoning loop.

AGAS pipeline


1. Repo overview

agent_attack_rs/
  bash/                           # Reviewer-friendly shell scripts (RQ1–RQ5)
  prompts/                        # Coordinator + per-role prompt templates
  scripts/run_agas.py             # Single entry point (--dataset --victim ...)
  src/agas/
    roles.py                      # Role enum {PR, SN, CA, IN} (paper symbols)
    signals.py                    # WorkerSignals (τ, γ, φ) + EnvSignals (ρ, Δρ, η, ξ, a)
    strategies.py                 # 8-strategy enum
    agents/                       # Coordinator + Worker policies
    simulation/                   # Episode runner (= AGAS algorithm outer loop)
    llm/                          # OpenAI / Ollama
    recsys/                       # Surrogate + 11-victim backends
    data/                         # Dataset loaders + preprocessing pipeline
  tests/                          # Pytest suite

2. Method recap

AGAS instantiates four worker roles (paper symbols in roles.py):

SymbolLong nameWhat it does
PRProfilerSafe filler-item ratings to probe the platform and build bridge pools.
SNSniperPayload role; direct target push or bridge-item promotion.
CACamouflageurStealth role; rebuilds trust with benign-looking activity.
INInactiveNo action this round (cool-down or quarantine).

Each round the Coordinator chooses exactly one of eight strategies from strategies.py (see method_strategies.tex):

  1. Victim Probe (S1_VICTIM_PROBE)
  2. Bridge Building (S2_BRIDGE_BUILDING, graph victims only)
  3. Warm-up (S3_WARM_UP)
  4. First Push (S4_FIRST_PUSH)
  5. Silent Slowdown (S5_SILENT_SLOWDOWN)
  6. Profile Cleanup (S6_PROFILE_CLEANUP)
  7. Safe Replacement (S7_SAFE_REPLACEMENT)
  8. Main Attack (S8_MAIN_ATTACK)

The Coordinator drives those decisions from two signal groups (signals.py):

  • Worker signals τ_{t,w}, γ_{t,w}, φ_{t,w} — trust, risk, and a structural validator. Update equations match method_coordinator.tex exactly (eq:trust_update, eq:risk_update, eq:risk_decay, eq:profile_validator).
  • Environment signals ρ^{(t)}, Δρ^{(t)}, η_t, ξ_t = (q_t, s_t), a_t — rank, rank-movement, acceptance rate, suppression signal, alert flag. The suspicion score q_t is the 0.2-weighted sum of the five normalised terms (d̂_t, δ̂_t, m̂_t, ŝ_t, g_t) from eq:round_suppression_terms and eq:round_suppression_score.

Round loop (ASCII)

                ┌─────────────────────────────────────────────────┐
   t=0…T-1 ──►  │ 1. Observe ρ^{(t)}, update memory m_t           │
                │ 2. Update τ, γ, φ, η, ξ, a                       │
                │ 3. Coordinator picks Strategy ∈ {S1…S8}          │
                │    and assigns Role ∈ {PR, SN, CA, IN} per worker│
                │ 4. Workers act (filler / bridge / target items)  │
                │ 5. Validate + accept actions → ΔR̃^{(t+1)}        │
                │ 6. Refit / query victim → ρ^{(t+1)}              │
                └─────────────────────────────────────────────────┘
                          │
                          ▼
                   t* = argmin_t ρ^{(t)}, return R* = [R ; R̃^{(≤t*)}]

The outer loop is implemented in src/agas/simulation/episode.py and mirrors algorithms/agas_end_to_end.tex.

3. Environment

ComponentRequirementTested with
Python≥ 3.103.13.5
PyTorch≥ 2.1 (targets only)2.11.0+cu128
CUDAoptional12.8
NumPy≥ 1.242.4.2
Pandas≥ 2.03.0.1
SciPy≥ 1.101.17.0
scikit-learn≥ 1.31.8.0
openai SDK≥ 1.122.21.0

PyTorch and CUDA are only required for the deep-learning victim models ([targets] extra). The core AGAS loop and the rule-based / surrogate paths run on CPU with no GPU dependency.

4. Install

pip install -e .
# If you want to use the deep-learning victim models (LightGCN, NeuMF, …)
pip install -e '.[targets]'

Required environment variables:

VariablePurposeDefault
OPENAI_API_KEYOpenAI Responses API key for the Coordinator / worker LLMs.(unset → fallback)
OPENAI_MODELModel name passed to OpenAI.gpt-5.1

4. Datasets

The paper evaluates on six public CF benchmarks (see experiment.tex):

Short nameSourceUsersItemsInteractionsDownloadPlace raw files in
ML-100KMovieLens 100K9431,682100,000GroupLensdata/ml-100k/
ML-1MMovieLens 1M6,0403,7061,000,209GroupLensdata/ml-1m/
Genome 2021MovieLens Tag Genome 202137,94184,6612,000,000 (capped)GroupLensdata/genome2021/
NetflixNetflix Prize342,44517,4342,000,000 (capped)Kaggledata/netflix/
DoubanDouban Movie28,05749,1768,085,679HKUSTdata/douban/
AmazonAmazon Reviews 2018998,65330,9642,000,000 (capped)UCSDdata/amazon/

After dropping the raw downloads into data/<dataset>/, run:

python scripts/preprocess_all.py --data-root data --output-root processed
Download Tool

Most Popular

View all →

Discover the most used tools by our community.

  • 1

    OsintHunter

    letox4/osinthunter

    🎯 The ultimate OSINT toolkit: investigation graph, confidence scoring, alias families, map, timeline and reports. 49 modules, 400+ sites, no API keys required.

    772
  • 2

    ghostlock-app

    yukonga/ghostlock-app

    GhostLock One-Tap Execution App (CVE-2026-43499)

    505
  • 3

    ARTEX

    baarai/artex

    LLM multi-agent autonomous penetration testing platform with planner/worker agents, asset and exploration graphs, MITM traffic recording, and vulnerability retesting.

    420
  • 4

    redforge

    cybersecurity91/redforge

    Browser-side red team toolkit with 102 offline tools for recon, payload generation, credential attacks, crypto, encoding and engagement reporting.

    345
  • 5

    CyberStrikeAI

    ed1s0nz/cyberstrikeai

    The system of action for AI-native cybersecurity—where intent becomes governed execution, evidence becomes operational memory, and every operation improves the next.

    260
  • 6

    antifox

    antifox-ai/antifox

    AI multi-agent autonomous penetration testing platform that runs recon, attack-surface discovery, exploitation validation and evidence-backed report delivery.

    251

Explore all tools

Browse our collection of tools

View all tools →