Distinguishing Attacks on Secure Channels usingMachine Learning

Abstract

Block ciphers are the most popular for protecting messages in the field of information security, and their power naturally draws attention. Identifying block ciphers in ECB and CBC mode has been difficult work over the past few decades. This paper proposes a completely new type of distinguishing attack in which we successfully generated ciphers from English text class and from random text class (i.e. rotated plaintext class) with circular rotation of plaintext bits, i.e., rotation by block size n, (first variation for n = 127) and (its other variation) rotation of plaintext (circular rotation) bits by length of plaintext (length (plaintext)-1), We encrypted plaintexts using DES, DES3, and AES in both variations, using ECB mode and CBC mode. Under ECB mode average accuracy for first and second variation is : 97.86% and 98.9% using Random forest by encryption using DES, 97% using SVM and 95% using Random forest by encryption using DES3, 86.15% and 91.3% using Random Forest by encryption using AES. Under CBC mode, average accuracy for first and second variation is : 52.30% & 52.59% using Logistic Regression by encryption using DES, 50.89% using Logistic Regression and 50.25% usingSVMby encryption using DES3, 50.8% using Logistic Regression and 50.1% using Random Forest by encryption using AES. The results demonstrate that ciphertext data can be successfully extracted by constructing a feature based on ciphertext recombination and location specificity.