K. S. Sesh Kumar

18 papers A 1B 1Journal 12Unranked 3
YearRankTypeTitle / Venue / Authors
2022 J jnl
J. Comput. Sci.
Caterina Buizza, César Quilodrán Casas, Philip Nadler, Julian Mack, Stefano Marrone, Zainab Titus, Clémence Le Cornec, Evelyn Heylen, Tolga Hasan Dur, Luis G. Baca Ruíz, Claire E. Heaney, Julio Amador Díaz López, K. S. Sesh Kumar, Rossella Arcucci
2022 J jnl
Trans. Mach. Learn. Res.
Michelangelo Conserva, Marc Peter Deisenroth, K. S. Sesh Kumar
2021 J jnl
CoRR
Samuel Cohen, K. S. Sesh Kumar, Marc Peter Deisenroth
2021 J jnl
CoRR
Michelangelo Conserva, Marc Peter Deisenroth, K. S. Sesh Kumar
2020 J jnl
Mach. Learn.
Riccardo Moriconi, Marc Peter Deisenroth, K. S. Sesh Kumar
2020 J jnl
Optim. Lett.
Riccardo Moriconi, K. S. Sesh Kumar, Marc Peter Deisenroth
2019 J jnl
CoRR
K. S. Sesh Kumar, Marc Peter Deisenroth
2019 J jnl
CoRR
Riccardo Moriconi, K. S. Sesh Kumar, Marc Peter Deisenroth
2017 J jnl
J. Mach. Learn. Res.
K. S. Sesh Kumar, Francis R. Bach
2016
K. S. Sesh Kumar
2015 J jnl
CoRR
K. S. Sesh Kumar, Álvaro Barbero Jiménez, Stefanie Jegelka, Suvrit Sra, Francis R. Bach
2013 conf
ICML (1)
K. S. Sesh Kumar, Francis R. Bach
2013 J jnl
CoRR
K. S. Sesh Kumar, Francis R. Bach
2012 J jnl
CoRR
K. S. Sesh Kumar, Francis R. Bach
2007 A conf
ICDAR
K. S. Sesh Kumar, S. Kumar, C. V. Jawahar
2006 B conf
Document Analysis Systems
Sachin Rawat, K. S. Sesh Kumar, Million Meshesha, Indraneel Deb Sikdar, A. Balasubramanian, C. V. Jawahar
2006 conf
ICVGIP
K. S. Sesh Kumar, Anoop M. Namboodiri, C. V. Jawahar
2005 conf
PReMI
K. S. Sesh Kumar, Anoop M. Namboodiri, C. V. Jawahar
redb/extractors/decompiler/bninja/similarity/minhasher.py
← Index redb/extractors/decompiler/bninja/similarity/minhasher.py python
import logging
import random
from enum import Enum

from ..analysis.medium_level_normalization import MediumLevelNormalization

try:
    from .minhashcustom import MinHashCustom
    from ..analysis.low_level_normalization import LowLevelNormalization
except ImportError:
    # Fallback to absolute imports (for multiprocessing spawned processes)
    from redb.extractors.decompiler.bninja.similarity.minhashcustom import MinHashCustom
    from redb.extractors.decompiler.bninja.analysis.low_level_normalization import LowLevelNormalization

## Values for this configuration were extracted from https://github.com/danielplohmann/mcrit/blob/main/mcrit/config/MinHashConfig.py#L10
# Length in number of Shingles of which a minhash consists
# this value represents the length of sha256sum hash truncated
MINHASH_SIGNATURE_LENGTH: int = 64
# Number of bits per signature element (1-32 bits)
MINHASH_SIGNATURE_BITS: int = 8


class TokenKind(Enum):
    LLIL = "llil"
    TYPED_LLIL = "typed_llil"
    MLIL = "mlil"
    TYPED_MLIL = "typed_mlil"


class MinHasher:
    # stick to the default method
    MINHASH_STRATEGY_HASH_ALL = 1

    def __init__(self, seed, il_function, kind: TokenKind = TokenKind.LLIL):
        self._minhash_seeds = []
        self.il_func = il_function
        self.kind = kind
        self._minhash_permutation = []
        self._signature_segments = []
        self._initMinhashing(seed)

    def _initMinhashing(self, MINHASH_SEED=None):
        random.seed(MINHASH_SEED)
        # init sequence of seeds
        self._minhash_seeds = [
            random.randint(0, MinHashCustom.getHashMax()) for _ in range(MINHASH_SIGNATURE_LENGTH)
        ]

    def make_ngrams(self, tokens, n=3):
        """Take the ngrams of the IL we try to pass into the functions"""
        return [tuple(tokens[i:i+n]) for i in range(len(tokens) - n + 1)]

    def _extract_tokens(self):
        """Extract the IL tokens from the IL function, picking the right
        normalizer (LLIL/MLIL) and the right normalization mode
        (skeleton/typed) based on self.kind."""
        if self.kind in (TokenKind.LLIL, TokenKind.TYPED_LLIL):
            normalizer = LowLevelNormalization()
        elif self.kind in (TokenKind.MLIL, TokenKind.TYPED_MLIL):
            normalizer = MediumLevelNormalization()
        else:
            raise ValueError(f"Unsupported token kind: {self.kind}")

        # typed variants include operand type info, skeleton variants don't
        if self.kind in (TokenKind.TYPED_LLIL, TokenKind.TYPED_MLIL):
            normalize = normalizer.normalize_instr_with_operands
        else:
            normalize = normalizer.normalize_instruction_all_levels

        instructions = []
        for basic_block in self.il_func.basic_blocks:
            for il in basic_block:
                instructions.append(normalize(il))

        return instructions

    def calculateMinHash(self):
        """Calculate hash function every time, then take minimum shingle per shingler"""
        minhash_result = MinHashCustom(minhash_bits=MINHASH_SIGNATURE_BITS)
        minhash_signature = []

        tokens = self._extract_tokens()
        shingles = self.make_ngrams(tokens, n=3)

        # Functions with fewer than 3 IL instructions can't produce n-grams
        # Return empty minhash for such small functions (thunks, stubs, etc.)
        # Triggered by 39d8ad95b0323c37bd3134ab93ac4af44c66a1a8443a41c1ac02cec19bb2816a
        if not shingles:
            return []

        # Generate the MinHash
        for seed in self._minhash_seeds:
            hashed_shingles = [
                self.shingle_hash(shingle, seed) for shingle in shingles
            ]
            min_value = min(hashed_shingles)

            if MINHASH_SIGNATURE_BITS < 32:
                min_value %= (2 ** MINHASH_SIGNATURE_BITS)

            minhash_signature.append(min_value)

        minhash_result.setMinHash(minhash_signature)
        return minhash_result.getMinHashInt()

    def shingle_hash(self, shingle, hash_seed=0):
        """Produce a single 32bit UINT hash for a given shingle"""
        return MinHashCustom.hashData(shingle, hash_seed)