RIPEMD-160 first-48-bit collision

Observed facts

The submitted inputs are distinct four-byte strings. Hex strings in collision.json have a 0x prefix; that prefix is notation, not part of the hashed input. Recomputing each full RIPEMD-160 digest with openssl dgst -ripemd160 -binary (OpenSSL 3.0.13) produced:

Input	Full 160-bit digest	First 48 bits
0x4b3fd300	a753ebe0c45aae16f469e0777f027348ee2bd8d0	a753ebe0c45a
0x942def00	a753ebe0c45ac7e92464fbe55b9f6a6e260405f5	a753ebe0c45a

The search hashed successive four-byte, little-endian encodings of integers starting at zero and found this pair after evaluating 15,674,773 inputs. A separate check through Python hashlib.new('ripemd160', data) agreed with both command-line digests. OpenSSL documents RIPEMD160 as a 160-bit digest and the dgst command's digest options.

Reproduction using the submitted file:

import hashlib, json
with open('collision.json', encoding='utf-8') as f:
    pair = json.load(f)
a, b = (bytes.fromhex(pair[k][2:]) for k in ('inputA', 'inputB'))
assert a != b
ha = hashlib.new('ripemd160', a).digest()
hb = hashlib.new('ripemd160', b).digest()
assert ha[:6] == hb[:6] == bytes.fromhex('a753ebe0c45a')
print(ha.hex(), hb.hex())

Inference

The two different inputs have the same first six digest bytes, which are the most significant 48 bits of the digest byte string. They therefore satisfy the requested truncated-digest collision for λ = 24.

Limits and unanswered questions

The full 160-bit digests differ. This is a collision only for the specified 48-bit truncation. The local checks use OpenSSL-backed paths; an independently implemented RIPEMD-160 library was not tested. SIMD's own recomputation remains the final external check.