Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Remote explainability faces the bouncer problem

Abstract

The concept of explainability is envisioned to satisfy society’s demands for transparency about machine learning decisions. The concept is simple: like humans, algorithms should explain the rationale behind their decisions so that their fairness can be assessed. Although this approach is promising in a local context (for example, the model creator explains it during debugging at the time of training), we argue that this reasoning cannot simply be transposed to a remote context, where a model trained by a service provider is only accessible to a user through a network and its application programming interface. This is problematic, as it constitutes precisely the target use case requiring transparency from a societal perspective. Through an analogy with a club bouncer (who may provide untruthful explanations upon customer rejection), we show that providing explanations cannot prevent a remote service from lying about the true reasons leading to its decisions. More precisely, we observe the impossibility of remote explainability for single explanations by constructing an attack on explanations that hides discriminatory features from the querying user. We provide an example implementation of this attack. We then show that the probability that an observer spots the attack, using several explanations for attempting to find incoherences, is low in practical settings. This undermines the very concept of remote explainability in general.

Access options

Rent or Buy article

Get time limited or full article access on ReadCube.

from$8.99

All prices are NET prices.

Fig. 1: Illustration of our model.
Fig. 2: The three scenarios involving remote explainability.
Fig. 3: Illustration of a possible implementation (Algorithm 1) of the PR attack.
Fig. 4: Percentage of label changes when swapping the discriminative features in the test set data for scenario B.
Fig. 5: Confidence level as a function of the number of tested input pairs, based on the German Credit detection probability in Fig. 4.
Fig. 6: Probability to find an IP, as a function of \({\mathbb{P}}(B)\), the probability of success for a non-discriminated group.

Data availability

The data that support the findings in this study—as the German Credit dataset—are publicly available at https://archive.ics.uci.edu/ml/datasets/statlog+(german+credit+data).

Code availability

The code used for the experiments is provided at https://github.com/erwanlemerrer/bouncer_problem (https://doi.org/10.5281/zenodo.3907271).

References

  1. 1.

    Veale, M. Logics and practices of transparency and opacity in real-world applications of public sector machine learning. In Proceedings of the 4th Workshop on Fairness, Accountability and Transparency in Machine Learning (FAT/ML, 2017); https://arxiv.org/pdf/1706.09249.pdf

  2. 2.

    de Laat, P. B. Algorithmic decision-making based on machine learning from big data: can transparency restore accountability? Philos. Technol. 31, 525–541 (2018).

    Article  Google Scholar 

  3. 3.

    Naumov, M., et al. Deep learning recommendation model for personalization and recommendation systems. Preprint at https://arxiv.org/pdf/1906.00091.pdf (2019).

  4. 4.

    Goodman, B. & Flaxman, S. European Union regulations on algorithmic decision-making and a ‘right to explanation’. AI Magazine 38, 50–57 (2017).

    Article  Google Scholar 

  5. 5.

    Selbst, A. D. & Powles, J. Meaningful information and the right to explanation. International Data Privacy Law 7, 233–242 (2017).

    Article  Google Scholar 

  6. 6.

    Adadi, A. & Berrada, M. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access 6, 52138–52160 (2018).

    Article  Google Scholar 

  7. 7.

    Guidotti, R. et al. A survey of methods for explaining black box models. ACM Comput. Surveys 51, 93 (2018).

    Google Scholar 

  8. 8.

    Molnar, C. Interpretable Machine Learning (GitHub, 2019); https://christophm.github.io/interpretable-ml-book/

  9. 9.

    Zhang, Y. & Chen, X. Explainable recommendation: a survey and new perspectives. Preprint at https://arxiv.org/pdf/1804.11192.pdf (2018).

  10. 10.

    Ribeiro, M. T., Singh, S. & Guestrin, C. ‘Why should I trust you?’: explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 1135–1144 (ACM, 2016); https://doi.org/10.1145/2939672.2939778

  11. 11.

    Galhotra, S., Brun, Y. & Meliou, A. Fairness testing: testing software for discrimination. In Proceedings of the 2017 11th Joint Meeting on Foundations of Software Engineering 498–510 (ESEC/FSE, 2017); https://doi.org/10.1145/3106237.3106277

  12. 12.

    Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems 4768–4777 (NIPS, 2017).

  13. 13.

    Andreou, A. et al. Investigating Ad Transparency Mechanisms in Social Media: A Case Study of Facebook’s Explanations (NDSS, 2018); https://doi.org/10.14722/ndss.2018.23204

  14. 14.

    Ateniese, G. et al. Provable data possession at untrusted stores. In Proceedings of the 14th ACM Conference on Computer and Communications Security 598–609 (ACM, 2007); https://doi.org/10.1145/1315245.1315318

  15. 15.

    Pearl, J. Causal inference in statistics: an overview. Stat. Surveys 3, 96–146 (2009).

    MathSciNet  Article  Google Scholar 

  16. 16.

    Aivodji, U. et al. Fairwashing: the risk of rationalization. In Proceedings of the 36th International Conference on Machine Learning (eds Chaudhuri, K. & Salakhutdinov, R.) 161–170 (PMLR, 2019).

  17. 17.

    Hajian, S., Domingo-Ferrer, J. & Martínez-Ballesté, A. Rule protection for indirect discrimination prevention in data mining. In Modeling Decision for Artificial Intelligence (eds Torra, V., Narakawa, Y., Yin, J. & Long, J.) 211–222 (Springer, 2011).

  18. 18.

    Menon, A. K. & Williamson, R. C. The cost of fairness in binary classification. In Proceedings of the 1st Conference on Fairness, Accountability and Transparency (eds Friedler, S. A. & Wilson, C.) 107–118 (PMLR, 2018).

  19. 19.

    Tramèr, F., Zhang, F., Juels, A., Reiter, M. K. & Ristenpart, T. Stealing machine learning models via prediction APIs. In Proceedings of the 25th USENIX Conference on Security Symposium, SEC’16 601–618 (USENIX Association, 2016).

  20. 20.

    Miller, T. Explanation in artificial intelligence: insights from the social sciences. Preprint at https://arxiv.org/pdf/1706.07269.pdf (2017).

  21. 21.

    Cummins, D. D., Lubart, T. & Alksnis, O. Conditional reasoning and causation. Memory Cognition 19, 274–282 (1991).

    Article  Google Scholar 

  22. 22.

    Alexander, L. What makes wrongful discrimination wrong? Biases, preferences, stereotypes and proxies. University of Pennsylvania Law Review 141, 149–219 (1992).

    Article  Google Scholar 

  23. 23.

    Wu, X. et al. Top 10 algorithms in data mining. Knowledge Inform. Syst. 14, 1–37 (2008).

    Article  Google Scholar 

  24. 24.

    Quinlan, J. R. C4.5: Programs for Machine Learning (Elsevier, 2014).

  25. 25.

    Statlog (German Credit Data) Data Set (UCI, accessed 1 September 2019); https://archive.ics.uci.edu/ml/datasets/Statlog+(German+Credit+Data)

  26. 26.

    Oreski, S. & Oreski, G. Genetic algorithm-based heuristic for feature selection in credit risk assessment. Expert Syst. Appl. 41, 2052–2064 (2014).

    Article  Google Scholar 

  27. 27.

    Brock, A., Donahue, J. & Simonyan, K. Large scale GAN training for high fidelity natural image synthesis. Preprint at https://arxiv.org/pdf/1809.11096.pdf (2019).

  28. 28.

    Khashman, A. Neural networks for credit risk evaluation: investigation of different neural models and learning schemes. Expert Syst. Appl. 37, 6233–6239 (2010).

    Article  Google Scholar 

  29. 29.

    Hou, J. et al. Ml defense: against prediction API threats in cloud-based machine learning service. In Proceedings of the International Symposium on Quality of Service, IWQoS ’19 7:1–7:10 (ACM, 2019)

  30. 30.

    Feldman, M., Friedler, S. A., Moeller, J., Scheidegger, C. & Venkatasubramanian, S. Certifying and removing disparate impact. In Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 259–268 (ACM, 2015); https://doi.org/10.1145/2783258.2783311

  31. 31.

    Braun, B. et al. Verifying computations with state. In Proceedings of the Twenty-Fourth ACM Symposium on Operating Systems Principles 341–357 (ACM, 2013); https://doi.org/10.1145/2517349.25227332013

  32. 32.

    Datta, A., Sen, S. & Zick, Y. Algorithmic transparency via quantitative input influence: theory and experiments with learning systems. In Proceedings of the 2016 IEEE Symposium on Security and Privacy (SP) 598–617 (IEEE, 2016).

  33. 33.

    Yeh, C.-K., Kim, J., Yen, I. E.-H. & Ravikumar, P. K. Representer point selection for explaining deep neural networks. In Proceedings of Advances in Neural Information Processing Systems 31 (eds Bengio, S. et al.) 9291–9301 (Curran Associates, 2018).

  34. 34.

    Tramèr, F., Zhang, F., Juels, A., Reiter, M. K. & Ristenpart, T. Stealing machine learning models via prediction APIs. In Proceedings of the 25th USENIX Security Symposium (USENIX Security 16) 601–618 (USENIX Association, 2016).

  35. 35.

    Milli, S., Schmidt, L., Dragan, A. D. & Hardt, M. Model reconstruction from model explanations. In Proceedings of the Conference on Fairness, Accountability and Transparency, FAT*19 1–9 (ACM, 2019).

  36. 36.

    Binns, R. Fairness in machine learning: lessons from political philosophy. In Proceedings of the 2018 Conference on Fairness, Accountability and Transparency Vol. 81, 149–159 (PMLR, 2017).

  37. 37.

    Mitchell, M. et al. Model cards for model reporting. In Proceedings of the Conference on Fairness, Accountability and Transparency, FAT*19 220–229 (ACM, 2019).

  38. 38.

    Blyth, C. R. On Simpson’s paradox and the sure-thing principle. J. Am. Stat. Assoc. 67, 364–366 (1972).

    MathSciNet  Article  Google Scholar 

  39. 39.

    Alipourfard, N., Fennell, P. G. & Lerman, K. Using Simpson’s paradox to discover interesting patterns in behavioral data. Preprint at https://arxiv.org/pdf/1805.03094.pdf (2018).

  40. 40.

    Zhang, L., Wu, Y. & Wu, X. Achieving non-discrimination in data release. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD17 1335–1344 (ACM, 2017).

  41. 41.

    Hajian, S. & Domingo-Ferrer, J. A methodology for direct and indirect discrimination prevention in data mining. IEEE Trans. Knowledge Data Eng. 25, 1445–1459 (2013).

    Article  Google Scholar 

  42. 42.

    Zhang, Y. & Zhou, L. Fairness assessment for artificial intelligence in financial industry. Preprint at https://arxiv.org/pdf/1912.07211.pdf (2019).

  43. 43.

    Tan, S., Caruana, R., Hooker, G. & Lou, Y. Distill-and-compare: auditing black-box models using transparent model distillation. In Proceedings of the 2018 AAAI/ACM Conference 303–310 AIES (AAAI, 2018); https://doi.org/10.1145/3278721.3278725

  44. 44.

    Chen, L., Mislove, A. & Wilson, C. Peeking beneath the hood of Uber. In Proceedings of the 2015 Internet Measurement Conference, IMC15 495–508 (ACM, 2015).

Download references

Author information

Affiliations

Authors

Contributions

The theoretical framework was developed by E.L.M. and G.T. Experimental work was carried out by E.L.M. and data analysis by G.T.

Corresponding authors

Correspondence to Erwan Le Merrer or Gilles Trédan.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Le Merrer, E., Trédan, G. Remote explainability faces the bouncer problem. Nat Mach Intell 2, 529–539 (2020). https://doi.org/10.1038/s42256-020-0216-z

Download citation

Search

Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing