Inspecting adversarial examples using the Fisher informationFθ=ED∼p(D∣θ)[∇θlogp(D∣θ)∇θ⊤logp(D∣θ)]=−ED∼p(D∣θ)[∇θ∇θ⊤logp(D∣θ)]\begin{aligned}\mathbb{F}_\theta&=\mathbb{E}_{\mathcal{D}\sim p(\mathcal{D}|\theta)}[\nabla_\theta \log p(\mathcal{D}|\theta
Inspecting adversarial examples using the Fisher informationFθ=ED∼p(D∣θ)[∇θlogp(D∣θ)∇θ⊤logp(D∣θ)]=−ED∼p(D∣θ)[∇θ∇θ⊤logp(D∣θ)]\begin{aligned}\mathbb{F}_\theta&=\mathbb{E}_{\mathcal{D}\sim p(\mathcal{D}|\theta)}[\nabla_\theta \log p(\mathcal{D}|\theta