Lessons from the AdKDD’21 Privacy-Preserving ML Challenge

Diemert, Eustache; Fabre, Romain; Gilotte, Alexandre; Jia, Fu; Leparmentier, Basile; Mary, Jérémie; Qu, Zhenzhen; Tanielian, Ugo; Yang, Hui

doi:10.1145/3485447.3512076

Cited by 2 publications

(7 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…where the quotient and multiplication are coordinate-wise. This "pseudo gradient" is very similar to the formula used in (Diemert et al, 2022) by the winners of the AdKdd challenge.…”

Section: Model Regularizationmentioning

confidence: 60%

“…Preprocessing on the Criteo-AdKdd challenge dataset This dataset is available in two versions: the noisy pre-aggregated data which were available to the challengers, and the full un-aggregated which was published with (Diemert et al, 2022). We found that our method was not performing well with the pre-aggregated data.…”

Section: Reported Metricmentioning

confidence: 97%

“…Criteo Ad-Kdd challenge Because learning a bidding model in this setting is a serious challenge for the advertising industry, Criteo proposed in 2021 a public competition where the goal was to learn a model from aggregated data (Diemert et al, 2022). We try to tackle a similar problem as the one from the competition, except that in the AdKKD's competition, a small amount of disaggregate examples were also provided and largely used by the top-performing solutions.…”

Section: Online Advertising Chrome Privacy Sandbox the Criteo-adkdd C...mentioning

confidence: 99%

“…• Criteo-AdKdd challenge (Diemert et al, 2022) is the largest dataset we used. The train set contains 80M samples, with 19 features each.…”

Section: Datasetsmentioning

confidence: 99%

“…Second, when re-aggregating, we were summing lots of instances of the privacy preserving noise (a Gaussian of std=17) together. While this noise is low enough to have a small impact when using directly the raw features, (see the experiments in (Diemert et al, 2022)) the variance of the summed noises was considerably larger, and it became an issue.…”

Section: Reported Metricmentioning

confidence: 99%

See 4 more Smart Citations

Learning from aggregated data with a maximum entropy model

Gilotte¹,

Yahmed²,

Rohde³

2022

Preprint

View full text Add to dashboard Cite

Aggregating a dataset, then injecting some noise, is a simple and common way to release differentially private data. However, aggregated data -even without noise-is not an appropriate input for machine learning classifiers. In this work, we show how a new model, similar to a logistic regression, may be learned from aggregated data only by approximating the unobserved feature distribution with a maximum entropy hypothesis. The resulting model is a Markov Random Field (MRF), and we detail how to apply, modify and scale a MRF training algorithm to our setting. Finally we present empirical evidence on several public datasets that the model learned this way can achieve performances comparable to those of a logistic model trained with the full unaggregated data.

show abstract

“…where the quotient and multiplication are coordinate-wise. This "pseudo gradient" is very similar to the formula used in (Diemert et al, 2022) by the winners of the AdKdd challenge.…”

Section: Model Regularizationmentioning

confidence: 60%

Section: Reported Metricmentioning

confidence: 97%

Section: Online Advertising Chrome Privacy Sandbox the Criteo-adkdd C...mentioning

confidence: 99%

“…• Criteo-AdKdd challenge (Diemert et al, 2022) is the largest dataset we used. The train set contains 80M samples, with 19 features each.…”

Section: Datasetsmentioning

confidence: 99%

Section: Reported Metricmentioning

confidence: 99%

See 3 more Smart Citations

Learning from aggregated data with a maximum entropy model

Gilotte¹,

Yahmed²,

Rohde³

2022

Preprint

View full text Add to dashboard Cite

show abstract

Who Ate All Our Cookies? Investigating Publishers’ Challenges Caused by Changes in Third-party Cookie Tracking

Stallone¹,

Gägauf²,

Kaya³

2022

Proceedings of the 18th International Conference on Web Information Systems and Technologies

View full text Add to dashboard Cite

This paper investigates the potential reactions of Swiss publishers, as actors with-in the digital advertising ecosystem, to the forthcoming fundamental changes to user tracking in the world wide web. The results of this mixed methodical study initiate the discussion on the future of cookie tracking by setting and then answering to four hypotheses regarding first-party tracking, shared ID solutions, Google's Privacy Sandbox, and a national walled garden system. The results show a clear inclination of Swiss publishers towards firstparty tracking and shared ID solutions, neutral standing towards Google's efforts to undo their harm provoking with their upcoming change, and an aversion towards a nation-wide walled garden. These findings intend to increase the volume of the discussion on the effects of BigTech's changes on the digital advertising ecosystem as a whole and therefore stimulate further research on the effects on single actors within this ecosystem -beyond the publishers themselves.

show abstract

Lessons from the AdKDD’21 Privacy-Preserving ML Challenge

Cited by 2 publications

References 10 publications

Learning from aggregated data with a maximum entropy model

Learning from aggregated data with a maximum entropy model

Who Ate All Our Cookies? Investigating Publishers’ Challenges Caused by Changes in Third-party Cookie Tracking

Contact Info

Product

Resources

About