TLDR: We suggest the uneven licensed robustness drawback, which requires licensed robustness for just one class and displays real-world adversarial situations. This targeted setting permits us to introduce feature-convex classifiers, which produce closed-form and deterministic licensed radii on the order of milliseconds.
Determine 1. Illustration of feature-convex classifiers and their certification for sensitive-class inputs. This structure composes a Lipschitz-continuous function map $varphi$ with a discovered convex perform $g$. Since $g$ is convex, it’s globally underapproximated by its tangent aircraft at $varphi(x)$, yielding licensed norm balls within the function house. Lipschitzness of $varphi$ then yields appropriately scaled certificates within the authentic enter house.
Regardless of their widespread utilization, deep studying classifiers are acutely weak to adversarial examples: small, human-imperceptible picture perturbations that idiot machine studying fashions into misclassifying the modified enter. This weak point severely undermines the reliability of safety-critical processes that incorporate machine studying. Many empirical defenses towards adversarial perturbations have been proposed—typically solely to be later defeated by stronger assault methods. We due to this fact deal with certifiably sturdy classifiers, which give a mathematical assure that their prediction will stay fixed for an $ell_p$-norm ball round an enter.
Typical licensed robustness strategies incur a variety of drawbacks, together with nondeterminism, sluggish execution, poor scaling, and certification towards just one assault norm. We argue that these points may be addressed by refining the licensed robustness drawback to be extra aligned with sensible adversarial settings.
The Uneven Licensed Robustness Downside
Present certifiably sturdy classifiers produce certificates for inputs belonging to any class. For a lot of real-world adversarial purposes, that is unnecessarily broad. Contemplate the illustrative case of somebody composing a phishing rip-off electronic mail whereas attempting to keep away from spam filters. This adversary will all the time try and idiot the spam filter into pondering that their spam electronic mail is benign—by no means conversely. In different phrases, the attacker is solely trying to induce false negatives from the classifier. Comparable settings embrace malware detection, faux information flagging, social media bot detection, medical insurance coverage claims filtering, monetary fraud detection, phishing web site detection, and lots of extra.
Determine 2. Uneven robustness in electronic mail filtering. Sensible adversarial settings typically require licensed robustness for just one class.
These purposes all contain a binary classification setting with one delicate class that an adversary is trying to keep away from (e.g., the “spam electronic mail” class). This motivates the issue of uneven licensed robustness, which goals to offer certifiably sturdy predictions for inputs within the delicate class whereas sustaining a excessive clear accuracy for all different inputs. We offer a extra formal drawback assertion in the primary textual content.
Characteristic-convex classifiers
We suggest feature-convex neural networks to deal with the uneven robustness drawback. This structure composes a easy Lipschitz-continuous function map ${varphi: mathbb{R}^d to mathbb{R}^q}$ with a discovered Enter-Convex Neural Community (ICNN) ${g: mathbb{R}^q to mathbb{R}}$ (Determine 1). ICNNs implement convexity from the enter to the output logit by composing ReLU nonlinearities with nonnegative weight matrices. Since a binary ICNN resolution area consists of a convex set and its complement, we add the precomposed function map $varphi$ to allow nonconvex resolution areas.
Characteristic-convex classifiers allow the quick computation of sensitive-class licensed radii for all $ell_p$-norms. Utilizing the truth that convex capabilities are globally underapproximated by any tangent aircraft, we are able to acquire a licensed radius within the intermediate function house. This radius is then propagated to the enter house by Lipschitzness. The uneven setting right here is vital, as this structure solely produces certificates for the positive-logit class $g(varphi(x)) > 0$.
The ensuing $ell_p$-norm licensed radius components is especially elegant:
[r*p(x) = frac{ color{blue}{g(varphi(x))} } { mathrm{Lip}_p(varphi) color{red}{| nabla g(varphi(x)) | *{p,*}}}.]
The non-constant phrases are simply interpretable: the radius scales proportionally to the classifier confidence and inversely to the classifier sensitivity. We consider these certificates throughout a variety of datasets, attaining aggressive $ell_1$ certificates and comparable $ell_2$ and $ell_{infty}$ certificates—regardless of different strategies usually tailoring for a particular norm and requiring orders of magnitude extra runtime.
Determine 3. Delicate class licensed radii on the CIFAR-10 cats vs canine dataset for the $ell_1$-norm. Runtimes on the appropriate are averaged over $ell_1$, $ell_2$, and $ell_{infty}$-radii (word the log scaling).
Our certificates maintain for any $ell_p$-norm and are closed kind and deterministic, requiring only one forwards and backwards go per enter. These are computable on the order of milliseconds and scale properly with community measurement. For comparability, present state-of-the-art strategies reminiscent of randomized smoothing and interval sure propagation usually take a number of seconds to certify even small networks. Randomized smoothing strategies are additionally inherently nondeterministic, with certificates that simply maintain with excessive likelihood.
Theoretical promise
Whereas preliminary outcomes are promising, our theoretical work suggests that there’s important untapped potential in ICNNs, even with no function map. Regardless of binary ICNNs being restricted to studying convex resolution areas, we show that there exists an ICNN that achieves excellent coaching accuracy on the CIFAR-10 cats-vs-dogs dataset.
Truth. There exists an input-convex classifier which achieves excellent coaching accuracy for the CIFAR-10 cats-versus-dogs dataset.
Nonetheless, our structure achieves simply $73.4%$ coaching accuracy with no function map. Whereas coaching efficiency doesn’t suggest take a look at set generalization, this consequence means that ICNNs are at the very least theoretically able to attaining the trendy machine studying paradigm of overfitting to the coaching dataset. We thus pose the next open drawback for the sphere.
Open drawback. Be taught an input-convex classifier which achieves excellent coaching accuracy for the CIFAR-10 cats-versus-dogs dataset.
Conclusion
We hope that the uneven robustness framework will encourage novel architectures that are certifiable on this extra targeted setting. Our feature-convex classifier is one such structure and offers quick, deterministic licensed radii for any $ell_p$-norm. We additionally pose the open drawback of overfitting the CIFAR-10 cats vs canine coaching dataset with an ICNN, which we present is theoretically potential. Please see our NeurIPS paper and codebase for extra particulars.