Phd Position F/m Auditing Online Ai Models

Le descriptif de l’offre ci-dessous est en Anglais_

**Type de contrat **:CDD

**Niveau de diplôme exigé **:Bac + 5 ou équivalent

**Fonction **:Doctorant

**A propos du centre ou de la direction fonctionnelle**:
The Inria center at the University of Rennes is one of eight Inria centers and has more than thirty research teams. The Inria center is a major and recognized player in the field of digital sciences. It is at the heart of a rich ecosystem of R&D and innovation, including highly innovative SMEs, large industrial groups, competitiveness clusters, research and higher education institutions, centers of excellence, and technological research institutes.

**Mission confiée**:
**Context**
The widespread use of black-box machine learning models in decision-making systems has
created a need for transparency and accountability. However, existing regulations and au
- diting techniques are insufficient to address the challenges posed by these opaque models.
These challenges include the need for efficient auditing (e.g. the computational tractability
of these audits), and the ability to handle continuously evolving models. While active au
- diting techniques show promise [3], they are limited to low-capacity models [1]. Thus, there
is a need for more efficient and scalable auditing techniques to address the challenges posed
by these modern black-box models. In addition, the auditing of remote online platforms’
models presents further challenges due to their constant evolution. Auditing algorithms
that can handle dynamic models and provide practical guidance for regulators are dearly
missing.

**Principales activités**:
**Research questions and objectives**
Although passive auditing offers a straightforward approach, its efficiency is limited. This
is because it treats each query independently, without considering the potential insights
gained from previous responses. This can lead to redundant queries and less efficient use
of the audit budget. Active auditing, on the other hand, leverages the information gained
from previous queries to select the most informative subsequent queries. This can signif
- icantly improve efficiency by focusing on areas of the input space that are most likely to
reveal insights about the model’s behavior. Active auditing offers potential advantages, but
also faces several challenges. It can be computationally expensive, especially for large-scale
models, limiting its practical applicability. Additionally, it often requires knowledge of the
model’s hypothesis space, which may not always be available or accurate. Finally, models
can be manipulated to mislead auditors [2], making it difficult to obtain accurate results.
To address the challenges faced by active auditing, future research could explore hybrid
approaches that combine passive and active auditing to balance efficiency and robustness.
More efficient active learning algorithms could be developed to reduce their computational
costs. Investigating methods to make active auditing more resistant to model manipulation
is another area of potential research. Furthermore, leveraging model explanations to guide
the selection of queries and improve the interpretability of audit results could also enhance
the effectiveness of auditing techniques for black-box models. By addressing these challenges
and exploring novel combinations, we plan to develop more effective and practical auditing.

Figure 1 provides an overview of the PACMAM project. On the left, the auditor is tasked
with auditing the target platform on the right. This platform exploits a model h, which
belongs to a given hypothesis space H. To perform her audit, the auditor sends requests
to the model (blue arrows) and gathers the corresponding model responses (red arrows).
The three red locks represent the identified scientific challenges: efficiency, tractability, and
dynamism. Furthermore, the concepts used to address these challenges are associated with
the work packages in which they intervene.
First, we will focus on a parameter that is paramount to audit difficulty: the capacity of
the audited model. While for the auditing of low capacity models the state-of-the-art has
shown that active approaches offer exponential budget improvements, we have shown in a
work entitled ”Under manipulations, are there AI models harder to audit?” [1] that when
auditing high capacity models, active and passive methods perform equally well in terms of
budget. Furthermore, we will investigate the situations between those two extreme cases,
namely by finely relating the target model capacity and its audit difficulty (defined as the
audit accuracy achievable under a fixed request budget).
We will then focus on devising tractable active auditing algorithms that exhibit an
acceptable computational complexity (supported by the auditor to find good inputs for
active auditing). Known active auditing algorithms are so computationally expensive that
their current practical applicabili