How kihy/anidsc works
Project type and purpose
• ANIDSC is an open-source Python toolkit for building and evaluating network intrusion detection pipelines, with support for adversarial attacks, streaming or batch data ingestion from pcap files, feature extraction, normalization, model training/inference (classical and deep-learning), drift detection, and result summarization/visualization.
• It is not a web app but a library plus reference scripts (adv_gen.py, live_detection.py, summarize_results.py) that runs in Docker or local Python environments.
File structure
• Top-level directories and scripts:
• docker_root (build files)
• src/ANIDSC (core library)
• features (BDD tests for chain behavior)
• test_data (example pcap files)
• deploy_commands, gen_uml.sh, git-delete (utility scripts)
• src/ANIDSC subdivides by function:
• data_source: offline_sources, live_sources
• component: pipeline primitives (pipeline_source, pipeline_component, processors, feature_buffer, feature_extractor, normalizer)
• pipeline: orchestration (Pipeline, multi_splitter)
• model: classical (boxplot, kitsune, gnnids), torch_model variants, node_encoder
• adversarial_attack: attack algorithms
• cdd_framework: drift detectors (arcus, drift_sense)
• evaluator: metrics, summarizer
• save_mixin: serialization backends
• utils, graph_rep, scaler, templates
Main components and relationships
a. Data Ingestion Layer
• offline_sources.live_sources → raw packet streams (pcap)
• offline_sources.offline_sources → historical batches
b. Processing & Feature Layer
• PipelineSource provides raw data to Pipeline
• PipelineComponent chain:
• FeatureBuffer (accumulates windows)
• FeatureExtractor modules (after_image, frequency, graph, synthetic, etc.)
• Normalizer / Scaler (t_digest)
c. Modeling Layer
• Models consume processed feature vectors:
• Classical detectors (boxplot, kitsune, gnnids)
• Deep-learning (PyTorch autoencoder, GOAD, ICL, SLAD, kitsune_torch)
• NodeEncoder (graph neural net input)
d. Drift Detection Layer
• CDD_Framework modules monitor concept drift on feature streams
e. Adversarial Attack Module
• Hooks into pipeline input or model gradient flow (liuer_mihou)
f. Evaluation & Reporting Layer
• Evaluator computes metrics (OD‐metrics), summarizer plots anomaly scores
• Summarize_results.py and graph_vis utilities for visual output
g. Persistence & Extensibility
• SaveMixin backends (json, yaml, pickle, torch, null) for model or pipeline checkpoints
• Templates and helper utilities for code generation
Architectural patterns and design principles
• Pipeline pattern: modular, pluggable components that adhere to a common interface
• Strategy pattern for feature extractors, models, save mixins
• Layered architecture: ingestion → processing → modeling → evaluation
• Separation of concerns: data sources, compute components, persistence, evaluation