Harden COCO dataset bootstrap: HTTPS + SHA-256 verification

[kiwaku]

Addresses the HTTPS + checksum-verification bullet of #2502.

Changes

• text_to_image/tools/coco.py, coco_generate_calibration.py, coco_calibration.py: upgrade the hardcoded COCO 2014 annotations URL from http://images.cocodataset.org/... to https://s3.amazonaws.com/images.cocodataset.org/... and verify the downloaded archive against a pinned SHA-256 before it is unzipped.
• coco.py and coco_calibration.py: per-image urllib.request.urlretrieve calls now rewrite http://images.cocodataset.org/ to the S3 origin so individual image downloads are also TLS-authenticated.
• Added a small _verify_sha256 helper and a COCO_ANNOTATIONS_TRAINVAL2014_SHA256 module constant to each of the three files.

Why route through s3.amazonaws.com directly instead of fixing the CDN hostname?

images.cocodataset.org terminates HTTPS with an S3 certificate whose subject is s3.amazonaws.com, so a naive http:// to https:// swap fails SNI certificate validation:

curl: (60) SSL: no alternative certificate subject name matches target host name 'images.cocodataset.org'

The path-style S3 URL (https://s3.amazonaws.com/images.cocodataset.org/...) serves the same bytes under a valid certificate, so we use that as the replacement origin. This preserves TLS authentication without depending on cocodataset.org fixing their CDN.

Checksum pinning

COCO_ANNOTATIONS_TRAINVAL2014_SHA256 is the SHA-256 of the canonical 252,872,794-byte archive (computed from the S3 origin at submission time). Mismatch raises RuntimeError before any unzip/read, so a tampered or truncated download cannot poison downstream calibration. If the archive is re-published and the hash rotates, maintainers can re-pin by running curl -sL <URL> | shasum -a 256 and updating the three constants.

Out of scope

• Other os.system calls in coco_calibration.py - those are tracked by the first bullet of Codebase refactor : Fix security and safety issues #2502 (partially addressed by merged Update coco.py to not utilize os.system #2489).
• Dataset URLs in files outside text_to_image/tools/ - did not scan the whole tree, happy to expand if maintainers want.

cc @tanvi-mlcommons (filed #2502), @arav-agarwal2 (driving the codebase refactor arc)

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[hanyunfan]

@manpreetssokhi Could you help to review it

[manpreetssokhi]

The links look okay to me in all 3 files, I don't think this change will cause any breaks and is okay to merge.

[kiwaku]

recheck

[hanyunfan]

@kiwaku Thanks for the contribution. Just want to let you know MLPerf Inference is a members‑only working group. Have you signed the CLA? If you’re already a member and the CLA is in place, I can help coordinate with someone to get you the necessary support.

[kiwaku]

@hanyunfan
Thank you, I’m not from a member organization.

I filled out the onboarding form, received the CLA by email, signed it, and sent it back ~6 days ago. The bot is still showing me as unsigned, so its probably not linked to my GitHub account yet.

I’d appreciate any help.

[hanyunfan]

@pgmpablo157321 Could you help to take a look at the CLA issue.