Wikipedia:Peer review/Differentially private stochastic gradient descent/archive1

Summarize Timeline Top Qs Fact Check

Article (edit | visual edit | history) · Article talk (edit | history) · Watch • Watch peer review

This peer review discussion is closed.

I've listed this article for peer review because this is my second article, and I would like to access my skills. I would like to push it to B-class, and I skipped AfC this time, so I have no clue if this is a decent article or not. Can someone please review this? Any suggestions would be appreciated. Thank you very much!

Sincerely,

16dvnk (talk) 01:06, 20 June 2026 (UTC)

There are two things I would point out.

The first is that the lead paragraph is much too long, I would add in breaks and refer to WP:LEADLENGTH.

Second, and much more importantly, there are only 7 citations supporting all of the article content. I haven’t read each citation, but my guess is there may potentially be some WP:SYNTH in the article that is not pure WP:WINOR. In general, using primary sources should only be for specific claims made in the source, especially regarding technical subjects, whereas secondary sources like review/survey papers can support broader generalizations about the field, and I think one (arguably two) of the papers in the article are review or survey papers. There are also a few sentences without citations, such as in Moments accountant, Applications and challenges, and also Research directions. I would also flag two of the seven citations being pre-prints from ArXiv, which generally causes skepticism on Wikipedia because they’re treated as self-published. These sourcing issues will be the biggest barrier to B-Class, and using survey/review papers for overview information, as well as using their bibliographies to look for more sources, might be a good approach.

Hope this helps! Jacob Rampino (talk) 00:52, 6 July 2026 (UTC)

Thank you very much for your review. I will apply edits shortly. 16dvnk (talk) 12:19, 6 July 2026 (UTC)

@Jacob Rampino Thanks for the review. I have tried to edit the article with the best of my ability:

• 1. Done. Copy editing and some rephrasing, but relatively easy.
• 2. a.) I didn't actually find any WP:SYNTH issues. Perhaps it is because the article is new, and this time I listened to WP:BACKWARDS except I only added a few more sources other than the base 3, since I was working on other articles concurrently. I added 3 new primary sources to support secondary sources. The article leans a little bit on primary sources though, but I am struggling to find new secondary sources. Hopefully it is enough, but if there are any issues (there probably are), we can work on it on the next round (since I need further guidance).
• b.) Identified the papers, they should be Ponomareva et al. and Monir et al. I did the necessary edits. There are still some primary sources used, but they should be for the straightforward facts and specific claims.
• c.) "There are also a few sentences without citations, such as in Moments accountant, Applications and challenges, and also Research directions." Unfortunately, I was only able to find 2 unsourced sentences in Research directions. I could not find any unsourced claims in the other sections. May you clarify? Sorry for this misunderstanding.
• d.) Luckily for me, the 2 preprints have peer reviewed versions. Swapped them out, no major changes needed for this point. (I addressed this before b.), so actually there is an edit) Monir et al. is a review paper, so I have to swap that out.

I believe I have addressed the peer review with the best of my ability. I will then scan the article for any copy edits. Thank you very much for your attention, and I appreciate you giving the time to this article.

Sincerely,

16dvnk (talk) 13:44, 7 July 2026 (UTC)

• My mistake on the ArXiv links! I missed the peer reviewed dois.
• As for the unsourced sentences, it's now clear you were citing at the end of paragraphs or long sections to avoid WP:CITEKILL, my mistake.

Overall this is a really fantastic article! Jacob Rampino (talk) 19:13, 8 July 2026 (UTC)

@Jacob Rampino Is this article B-class ready? Do you think this article has potential for GA nomination in the near future (a quick review on the B/GA criteria would be very great for me to improve it further)? Thank you very much for your time and reviews.

Sincerely,

16dvnk (talk) 12:03, 9 July 2026 (UTC)

I would say it is certainly C-class at least. For the B-Class criteria:

• Referencing and citations: Y
• Coverage and accuracy: N
• Structure: Y
• Grammar: Y
• Supporting materials: N
• Accessibility: N

The citations are reliable and used throughout.

Coverage and accuracy doesn’t meet B-class because the article omits tooling (e.g. Opacus, TensorFlow Privacy, JAX-Privacy), ε/δ examples, and a clear technical breakdown of the algorithm itself, such as structured pseudocode. The low diversity of sources doesn’t give a sense of the wider DP-SGD literature. This is especially true of applications, the "Applications and challenges" section names relevant domains like healthcare and cybersecurity, but mostly repeats the lead rather than going through specific examples within those domains.

Grammar and structure are robust.

Supporting materials fails B-class because the article has no images, diagrams, tables, or pseudocode at all. This is a step by step algorithm involving gradient clipping and noise addition, so a diagram or algorithm box would normally be expected to help readers follow the process.

Accessibility fails B-class because the article relies on specific technical terminology, like Rényi divergence, moments accountant, L2 norm clipping, etc., without plain-language explanation that would help a non-specialist reader. Several sections assume familiarity with differential privacy and optimization concepts, which makes the article harder to follow for readers outside the field. This criterion is usually the hardest to achieve for technical articles like this.

Sorry for the delayed response! Jacob Rampino (talk) 16:05, 19 July 2026 (UTC)

Alright, I will start addressing these issues. Thank you very much! 16dvnk (talk) 23:29, 19 July 2026 (UTC)

(1/n) Alright, I have finally stepped aside from a complex dispute, and I can start addressing these issues. Special thanks to the extensive review. I will first address the coverage issues.

Details:

• Coverage and accuracy: This part makes sense; I did minimal research and only started with a narrow selection of sources. Will include Opacus etc... Edit: this part is addressed

• Structured pseudocode: Turns out, the Abadi et al. paper already has one, where I have overlooked. Added it in.

• Opacus: Added one small paragraph that should be due-weight, but if not, I can expand it in the future, but I could not find many reliable secondary sources about it.

• TensorFlow Privacy: This was harder to find, but eventually I have finally found a paper about it. Done.

• JAX-Privacy: Similar to Opacus and TensorFlow Privacy. Done. I have also found an additional bridging source stating that these 3 methods are "common", which is useful to the article.

• Applications:

I have found 2 applications, including healthcare applications and cybersecurity applications:

• Healthcare applications: A peer reviewed paper about clinical strokes, with a 93% accuracy. I found it useful and I added it in.

• Cybersecurity applications: Another peer reviewed paper caught my eye. Not much to say here. Done. Lead: Updated to reflect everything.

With all these changes It seems like the coverage issues are mostly addressed. Maybe some minor polish is needed, but I will go through those in the next cycle. Fixes for other issues are pending. 16dvnk (talk) 03:57, 23 July 2026 (UTC)

P.S. I might have to move my response here to the PR? Not sure how this works, since I am relatively new to this process. 16dvnk (talk) 04:07, 23 July 2026 (UTC)

(2/n) Supporting materials:

My previous articles did not have any supporting materials at all, because it was unnecessary. This is relatively new to me, but I will try my best here.

• Pseudocode: Added. I tried my best to follow the MOS for this, but I may stand corrected since I have no profession in writing pseudocode.
• Images: I initially created an svg, but then I decided a table would be clearer.
• Tables: Included in the Description. I do not see any other place to add one. Done.

The support materials should be addressed, but I have to admit I am new with this. I may stand corrected. The Accessibility issues will be addressed next turn. Thank you for your attention. 16dvnk (talk) 04:49, 23 July 2026 (UTC)

(3/3) Accessibility:

This should be relatively easy. Let me pinpoint some areas:

• Moments accountant (and RDP): Defined what a moments accountant is. Then, I also defined what kurtosis and skewness is (which probably does not need a source according to WP:OR). Also defined RDP, with explaining what an RD is.
• Description: Defined gradient and mini-batch (which does not need a source with the same reasons as above). Also defined subsampling etc. All terms are explained.
• Lead: Defined terms like non-convex, privacy budgets, gradients, etc.
• Software implementations: Explained optimizer.
• Applications and challenges: Explained F1-score, memory overhead, etc.
• Research directions: Explained lr, budget allocation and empirical privacy auditing.
• In the code: Explained loss function. Updated from "text" to "wikitext" to make the bolding work.
• Fixed some typos

With this copy edit, it seems like it is more accessible that before. The only question is if we need citations for some of the definitions, or if we should put those in footnotes, but other than that, there appears to be no major term that users may not understand. I may stand corrected.

With that, it seems like I am done with this round of peer review, as I do not find any other glaring issues. Thank you for your extensive review!

Sincerely,

16dvnk (talk) 06:50, 23 July 2026 (UTC)