Energy-Based Cipher Classifier

Purpose

The energy classifier is used to classify cipher-related patterns and distinguish between different cipher behaviours present in the project. Instead of directly predicting the secret key, the model learns to identify whether a given feature vector corresponds to a valid cipher transformation.

This module acts as the first stage of the pipeline by learning statistical representations of different cipher types and their structural behaviour.

Cipher Types

The project includes the following cipher categories:

  • Advanced Encryption Standards(AES)
  • Classical substitution-based cipher (Vigenère)
  • Lightweight block cipher (SPECK32)
  • Feistel-based block cipher (DES)

Each cipher exhibits distinct statistical and structural patterns. The energy classifier learns to differentiate these patterns.

Theoretical Foundation

An Energy-Based Model defines a scalar function:

E(x, y)

where:

x = input feature vector (trace-derived or text-derived features)
y = cipher class label

The model assigns:

Low energy → correct cipher class
High energy → incorrect cipher class

Classification Mechanism

During training, the model learns:

  • statistical structure of ciphertext or trace features
  • distributional differences between cipher types
  • structural properties of transformations

This enables the classifier to distinguish between:

  • classical polyalphabetic behaviour
  • Feistel network leakage patterns
  • ARX-based lightweight cipher leakage

Training Objective

The loss function enforces:

E(x, y_true) < E(x, y_false)

This creates an energy margin between correct and incorrect cipher classes.

Output Interpretation

The model produces an energy score for each cipher type.
The predicted class is the one with minimum energy.

This allows:

  • multi-class cipher classification
  • statistical validation of cipher behaviour
  • feature learning for downstream cryptanalysis

Accuracy

alt text

Role in the Cryptanalysis Pipeline

The energy classifier:

  • learns structural differences between cipher families
  • provides feature-level understanding of cipher behaviour
  • acts as a preprocessing and analysis stage before key recovery attacks