Computer Vision MT25, Ethics and privacy

Other features may correlate with the sensitive features.

The classifier response is independent from the sensitive attribute, i.e.

The classifier response is conditionally independent from the sensitive attribute given the target, i.e.

All groups (delineated by $A$) experience the same false negative and false positive rate.

Proof: Separation says $R$ is conditionally independent of $A$ given $Y$. Combining the stated factorisation with the chain rule $\mathbb P(R, A \mid Y) = \mathbb P(R \mid Y, A)\,\mathbb P(A \mid Y)$ and cancelling $\mathbb P(A \mid Y) > 0$ gives

so the prediction depends on the target alone, not the group. Reading off the two error rates for any group $A = a$ (binary $Y, R$):

• FPR: $\mathbb P(R = 1 \mid Y = 0, A = a) = \mathbb P(R = 1 \mid Y = 0)$,
• FNR: $\mathbb P(R = 0 \mid Y = 1, A = a) = \mathbb P(R = 0 \mid Y = 1)$.

Both right-hand sides are independent of $a$, so every group shares the same FPR and the same FNR. $\square$

Suppose further:

• $Y$ is binary
• $A$ is not independent of $Y$
• $R$ is not independent of $Y$

Then:

• Both independence and separation cannot hold simultaneously.

Proof: Suppose both hold. Write $r _ y := \mathbb P(R = 1 \mid Y = y)$ and $q _ a := \mathbb P(Y = 1 \mid A = a)$. Condition $R = 1$ on group $a$ and expand over the binary $Y$:

where separation let us replace $\mathbb P(R = 1 \mid Y = y, A = a)$ by $r _ y$ (free of $a$), and $Y$ binary gave $\mathbb P(Y = 0 \mid A = a) = 1 - q _ a$.

By independence the left side is constant in $a$, so $(r _ 1 - r _ 0)\, q _ a$ is constant in $a$. But $A \not\perp Y$ with $Y$ binary means $q _ a$ is not constant in $a$, forcing $r _ 1 = r _ 0$, i.e.

contradicting $R \not\perp Y$. (Running the same argument for each value of $R$ removes the binary-$R$ assumption.) $\square$

Harms caused by a system allocating resources unfairly.

Where a system reinforces harmful stereotypes.

• Recognition: A group is erased or made invisible.
• Under-representation: A group is under-represented.
• Denigration: Use of culturally disparaging terms.
• Ex-nomination: Representing ideology as common sense.
• Stereotype: Reinforces negative stereotypes.

(Mnemonic: RUDES).

Recall the gender bias of a category $C$:

Bias amplification is the observation that the predicted gender ratio on the test data can actually exceed the true training gender ratio (perhaps because accuracy-maximisation on imbalanced data optimally predicts the majority, or the model picks up shortcut features).

• Defendants care about False Positive Rate $= \mathrm{FP} / (\mathrm{FP} + \mathrm{TN})$, the rate at which non-offenders are wrongly predicted to reoffend. A high FPR means innocent-of-reoffence defendants get long sentences.
• Judges (and the legal system more broadly) care about False Negative Rate $= \mathrm{FN} / (\mathrm{FN} + \mathrm{TP})$, the rate at which true reoffenders are wrongly predicted to be low-risk. A high FNR means released defendants commit more crimes.

These two stakeholder concerns are in tension — driving FPR down (favouring defendants) typically raises FNR (worsening for judges), and vice versa.

• Datasheets for Datasets (Gebru et al., FAccT 2018): a standard list of questions to answer when releasing a dataset — who created it, why, what is in it, who labelled the images and how, etc. Aim is to give downstream users enough context to assess fit-for-purpose.
• Model Cards (Mitchell et al., FAccT 2019): a standard list of questions to answer when releasing a trained model — intended use, factors, metrics, evaluation/training data, quantitative analyses, ethical considerations, caveats. Importantly, Model Cards include explicit out-of-scope use cases (e.g. the CLIP model card states “any deployed use case is currently out of scope”).

Both have been adopted in practice by Google and (sometimes) OpenAI.

CelebA (Liu et al., ICCV 2015) tags each of 202k face images with 40 binary attributes. Many of these attributes are deeply subjective: Attractive, Big _ Lips, Big _ Nose, Chubby, Heavy _ Makeup, High _ Cheekbones, Narrow _ Eyes, Oval _ Face, Pointy _ Nose, Wavy _ Hair, Young — and, notably, Male (treated as binary).

The critique asks: who chose these attributes? Why these and not others? How are they operationally defined? Who actually labelled the images? The paper’s answers are very thin: “annotated by a professional labeling company”. No annotator guidelines, no inter-annotator agreement, no demographic breakdown of labellers. Yet thousands of downstream papers train and evaluate face models against these labels as if they were objective.

This is the standard exam-relevant example of subjective labelling masquerading as ground truth.

Drawing a coloured annotation (e.g. a red circle) on an image steers a VLM’s global descriptor toward the annotated region. The lecture’s worked example: an image of a camouflaged owl on a tree trunk.

• Without annotation: CLIP classifies the image as “tree” with confidence $\sim 0.443$ and assigns negligible mass to “owl”.
• With a red circle drawn around the owl: CLIP’s confidence for “owl” rises to $\sim 0.925$.

Caveat from the lecture: the same trick steers the model toward whatever the circle highlights, including stereotypes — the same setup that fixes the owl can also bias a model toward labelling a circled face as “missing person” or “murderer”. So the technique is powerful but also reveals training-data biases.