On Fairness of Low-Rank Adaptation of Large Models

We thank the reviewer for their time and are glad that the reviewer finds our work valuable! We respond to the main comments below:

My own observation is that there is no significant difference of fairness metrics between LoRA fine-tuned and fully fine-tuned models. I don't think such observation implies anything on whether LoRA may exacerbate the biases or not, given the fact that the learning objectives are still identical.

We agree with the reviewer that our study provides no evidence for LoRA exacerbating unfairness (i.e. our study, albeit comprehensive, can possibly, but unlikely, be a false negative) and this is indeed distinct from having evidence that LoRA does not exacerbate fairness. We have acknowledged this limitation in Sec 4 and will ensure this is clear in the updated version.

Is the model fine-tuned subject to fairness constraints?

No, both LoRA and full fine-tuning are performed with standard Adam. We agree that it would be interesting to see how fair training (e.g. fairness constraints, or fair learning algorithms) will influence the fairness results. However, as we discussed at the end of section 2, since our end goal is to compare LoRA with full fine-tuning, we believe that the fine-tuning task should not teach the model to be fair, since it would otherwise introduce confounding factors of whether LoRA or full FT responds to a particular fairness constraint or fair learning algorithm better. Teaching the model to be fair would make whatever fairness conclusions we draw less generalizable to unseen tasks and models.

Nevertheless, the reviewer has raised a good point and we hope to add relevant discussions in the updated version.

Thank you for raising interesting questions! However, the reviewer’s comments suggest a mischaracterization of our work and we hope to clarify below:

Let's use a metaphor

The logic of the metaphor is as follows:

1. LoRA acts indiscriminately on “all the model parameters”;
2. Hence, its side effects (if any) are also applied indiscriminately to all aspects of model behaviors, not just fairness (“blurs the whole picture”);
3. Hence, why do the authors (only) study fairness and why “this low-ranking should act only on fairness”?

Unfortunately, this metaphor has critical flaws.

1. While true that LoRA is broadly applicable, it is unclear from prior work that “it has effects on every aspect of LMs”: many aspects like robustness (OOD or adversarial), privacy, and fairness are unexplored against full fine-tuning, and our work is one of the first to provide such a study.
2. We’re unaware of evidence for the reviewer’s claim that LoRA “has effects on every aspect of LMs” just because it applies to “all the parameters”.
3. Even if LoRA may have “effects on every aspect of LMs”, it does not mean that “this low-ranking should act only on fairness” and we made no such claim.

Provide a good reason why LoRA should have any influence on fairness

(To avoid duplication, please also see our response to reviewer yDCp.)

We emphasize that our main motivation is precisely that the impact on fairness is unclear from prior work. Our study was also in part motivated by the disparate impact of pruning and differential privacy on subgroup fairness.

Why focus on fairness? Is there any reason why this low-ranking should act only on fairness?

There are indeed many aspects that we could focus on. However, fairness is an important practical aspect deserving its own investigation (as opposed to spreading our efforts to more aspects but being less comprehensive).

Indeed, many other aspects may be impacted by LoRA, and that the fairness impact could be an indirect result these other aspects. While true, this does not mean that only fairness is impacted or worth studying, or “this low-ranking should act only on fairness”. We have made no such claims. We added LoRA and observed the actual fairness metrics end-to-end; thus while it is possible to have other mechanisms at play, our study still provides useful insight into how LoRA behaves in the real world.

Nevertheless, we agree that it is interesting future work to study other potential (side) effects of LoRA.

Thank you for the thoughtful review! We respond to the main points below:

The paper's premise is unclear, why should LORA worsen the fairness of the model decisions?

• We’d like to clarify that we in fact did not begin with a premise that LoRA would worsen fairness (indeed, it may even improve it as seen in Sec 3.2). Rather, we believe the impact on fairness is unclear from prior studies. This lack of clarity, together with LoRA’s wide adoption, may imply unintended and harmful consequences in high-stakes applications and has motivated our thorough investigation.
• Our study of LoRA fairness was also motivated by the disparate impact of pruning [1] and differential privacy [2] on subgroup fairness. A common thread is that both pruning and DP reduce a model’s fitting capacity, which [3] argues can lead to disproportionate effects on minority data. Since LoRA also sacrifices fitting capacity for efficiency, it is natural to ask whether a similar impact exists.

Shouldn't it just be a function of the underlying data that is being used? With this logic, shouldn't the experimental results be expected (or even obvious)?

• We respectfully disagree that fairness is always solely a simple function of data; indeed, DP training is a counter-example where the learning algorithm impacts fairness, despite the same data. We also visualized the fairness impact of subgroup sizes in Appendix D.5, where subgroups with fewer examples (a common factor believed to tie to fairness) do not necessarily experience worse fairness.
• Generally, expecting results does not imply that such results were previously known. We believe a thorough study confirming (or disproving) intuitions is useful to the community.

study the impact of other PEFT methods like adaptors, IA3, etc.

• We aim to be comprehensive on one popular method rather than to spread efforts across multiple methods. Indeed, recent work such as https://arxiv.org/pdf/2405.09673 suggests that LoRA’s wide-spread use warrants a focused study.
• Nevertheless, we acknowledge this important future work in concluding remarks.

We hope to make these discussions clearer in the updated version!

[1] Tran et al., “Pruning has a disparate impact on model accuracy.” NeurIPS 2022.

[2] Bagdasaryan et al., "Differential privacy has disparate impact on model accuracy." NeurIPS 2019.

[3] Feldman & Zhang, "What neural networks memorize and why: Discovering the long tail via influence estimation." NeurIPS 2020.

We thank the reviewer for their time and are glad that the reviewer finds our work valuable! Trustworthiness aspects (such as fairness) of popular algorithms are often overlooked by practitioners and we believe a thorough empirical investigation can help inform the community and future applications.