FEATURE: Classify posts looking for NSFW images

This commit is contained in:
Roman Rizzi 2023-02-24 09:11:58 -03:00
parent 94933f3c58
commit 85768cfb1c
No known key found for this signature in database
GPG Key ID: 64024A71CE7330D3
8 changed files with 146 additions and 53 deletions

View File

@ -11,7 +11,7 @@ module DiscourseAI
def inject_into(plugin) def inject_into(plugin)
nsfw_detection_cb = nsfw_detection_cb =
Proc.new do |post| Proc.new do |post|
if SiteSetting.ai_nsfw_detection_enabled if SiteSetting.ai_nsfw_detection_enabled && post.uploads.present?
Jobs.enqueue(:evaluate_post_uploads, post_id: post.id) Jobs.enqueue(:evaluate_post_uploads, post_id: post.id)
end end
end end

View File

@ -3,15 +3,19 @@
module Jobs module Jobs
class EvaluatePostUploads < ::Jobs::Base class EvaluatePostUploads < ::Jobs::Base
def execute(args) def execute(args)
upload = Upload.find_by_id(args[:upload_id]) return unless SiteSetting.ai_nsfw_detection_enabled
return if (post_id = args[:post_id]).blank?
return unless upload post = Post.includes(:uploads).find_by_id(post_id)
return if post.nil? || post.uploads.empty?
result = DiscourseAI::NSFW::Evaluation.new.perform(upload) nsfw_evaluation = DiscourseAI::NSFW::Evaluation.new
# FIXME(roman): This is a simplistic action just to create image_uploads = post.uploads.select { |upload| FileHelper.is_supported_image?(upload.url) }
# the basic flow. We'll introduce flagging capabilities in the future.
upload.destroy! if result[:verdict] results = image_uploads.map { |upload| nsfw_evaluation.perform(upload) }
DiscourseAI::FlagManager.new(post).flag! if results.any? { |r| r[:verdict] }
end end
end end
end end

View File

@ -8,7 +8,7 @@ module ::Jobs
post_id = args[:post_id] post_id = args[:post_id]
return if post_id.blank? return if post_id.blank?
post = Post.includes(:user).find_by(id: post_id, post_type: Post.types[:regular]) post = Post.find_by(id: post_id, post_type: Post.types[:regular])
return if post&.raw.blank? return if post&.raw.blank?
::DiscourseAI::Toxicity::PostClassifier.new.classify!(post) ::DiscourseAI::Toxicity::PostClassifier.new.classify!(post)

View File

@ -13,7 +13,14 @@ module ::DiscourseAI
end end
def flag! def flag!
PostActionCreator.create(@flagger, @object, :inappropriate, reason: @reasons) PostActionCreator.new(
@flagger,
@object,
PostActionType.types[:inappropriate],
reason: @reasons,
queue_for_review: true,
).perform
@object.publish_change_to_clients! :acted @object.publish_change_to_clients! :acted
end end
end end

View File

@ -6,18 +6,17 @@ describe DiscourseAI::NSFW::EntryPoint do
fab!(:user) { Fabricate(:user) } fab!(:user) { Fabricate(:user) }
describe "registering event callbacks" do describe "registering event callbacks" do
fab!(:image_upload) { Fabricate(:upload) }
let(:raw_with_upload) { "A public post with an image.\n![](#{image_upload.short_path})" }
before { SiteSetting.ai_nsfw_detection_enabled = true }
context "when creating a post" do context "when creating a post" do
let(:creator) do let(:creator) do
PostCreator.new( PostCreator.new(user, raw: raw_with_upload, title: "this is my new topic title")
user,
raw: "this is the new content for my topic",
title: "this is my new topic title",
)
end end
it "queues a job on create if sentiment analysis is enabled" do it "queues a job on create if sentiment analysis is enabled" do
SiteSetting.ai_nsfw_detection_enabled = true
expect { creator.create }.to change(Jobs::EvaluatePostUploads.jobs, :size).by(1) expect { creator.create }.to change(Jobs::EvaluatePostUploads.jobs, :size).by(1)
end end
@ -26,6 +25,13 @@ describe DiscourseAI::NSFW::EntryPoint do
expect { creator.create }.not_to change(Jobs::EvaluatePostUploads.jobs, :size) expect { creator.create }.not_to change(Jobs::EvaluatePostUploads.jobs, :size)
end end
it "does nothing if the post has no uploads" do
creator_2 =
PostCreator.new(user, raw: "this is a test", title: "this is my new topic title")
expect { creator_2.create }.not_to change(Jobs::EvaluatePostUploads.jobs, :size)
end
end end
context "when editing a post" do context "when editing a post" do
@ -33,9 +39,7 @@ describe DiscourseAI::NSFW::EntryPoint do
let(:revisor) { PostRevisor.new(post) } let(:revisor) { PostRevisor.new(post) }
it "queues a job on update if sentiment analysis is enabled" do it "queues a job on update if sentiment analysis is enabled" do
SiteSetting.ai_nsfw_detection_enabled = true expect { revisor.revise!(user, raw: raw_with_upload) }.to change(
expect { revisor.revise!(user, raw: "This is my new test") }.to change(
Jobs::EvaluatePostUploads.jobs, Jobs::EvaluatePostUploads.jobs,
:size, :size,
).by(1) ).by(1)
@ -44,7 +48,14 @@ describe DiscourseAI::NSFW::EntryPoint do
it "does nothing if sentiment analysis is disabled" do it "does nothing if sentiment analysis is disabled" do
SiteSetting.ai_nsfw_detection_enabled = false SiteSetting.ai_nsfw_detection_enabled = false
expect { revisor.revise!(user, raw: "This is my new test") }.not_to change( expect { revisor.revise!(user, raw: raw_with_upload) }.not_to change(
Jobs::EvaluatePostUploads.jobs,
:size,
)
end
it "does nothing if the new raw has no uploads" do
expect { revisor.revise!(user, raw: "this is a test") }.not_to change(
Jobs::EvaluatePostUploads.jobs, Jobs::EvaluatePostUploads.jobs,
:size, :size,
) )

View File

@ -1,32 +0,0 @@
# frozen_string_literal: true
require "rails_helper"
require_relative "../../../../../support/nsfw_inference_stubs"
describe Jobs::EvaluatePostUploads do
fab!(:image) { Fabricate(:s3_image_upload) }
describe "#execute" do
before { SiteSetting.ai_nsfw_inference_service_api_endpoint = "http://test.com" }
context "when we conclude content is NSFW" do
before { NSFWInferenceStubs.positive(image) }
it "deletes the upload" do
subject.execute(upload_id: image.id)
expect { image.reload }.to raise_error(ActiveRecord::RecordNotFound)
end
end
context "when we conclude content is not NSFW" do
before { NSFWInferenceStubs.negative(image) }
it "does nothing" do
subject.execute(upload_id: image.id)
expect(image.reload).to be_present
end
end
end
end

View File

@ -0,0 +1,100 @@
# frozen_string_literal: true
require "rails_helper"
require_relative "../../../../../support/nsfw_inference_stubs"
describe Jobs::EvaluatePostUploads do
describe "#execute" do
before do
SiteSetting.ai_nsfw_detection_enabled = true
SiteSetting.ai_nsfw_inference_service_api_endpoint = "http://test.com"
end
fab!(:upload_1) { Fabricate(:s3_image_upload) }
fab!(:post) { Fabricate(:post, uploads: [upload_1]) }
describe "scenarios where we return early without doing anything" do
before { NSFWInferenceStubs.positive(upload_1) }
it "does nothing when ai_toxicity_enabled is disabled" do
SiteSetting.ai_nsfw_detection_enabled = false
subject.execute({ post_id: post.id })
expect(ReviewableFlaggedPost.where(target: post).count).to be_zero
end
it "does nothing if there's no arg called post_id" do
subject.execute({})
expect(ReviewableFlaggedPost.where(target: post).count).to be_zero
end
it "does nothing if no post match the given id" do
subject.execute({ post_id: nil })
expect(ReviewableFlaggedPost.where(target: post).count).to be_zero
end
it "does nothing if the post has no uploads" do
post_no_uploads = Fabricate(:post)
subject.execute({ post_id: post_no_uploads.id })
expect(ReviewableFlaggedPost.where(target: post_no_uploads).count).to be_zero
end
it "does nothing if the upload is not an image" do
SiteSetting.authorized_extensions = "pdf"
upload_1.update!(original_filename: "test.pdf", url: "test.pdf")
subject.execute({ post_id: post.id })
expect(ReviewableFlaggedPost.where(target: post).count).to be_zero
end
end
context "when the post has one upload" do
context "when we conclude content is NSFW" do
before { NSFWInferenceStubs.positive(upload_1) }
it "flags and hides the post" do
subject.execute({ post_id: post.id })
expect(ReviewableFlaggedPost.where(target: post).count).to eq(1)
expect(post.reload.hidden?).to eq(true)
end
end
context "when we conclude content is safe" do
before { NSFWInferenceStubs.negative(upload_1) }
it "does nothing" do
subject.execute({ post_id: post.id })
expect(ReviewableFlaggedPost.where(target: post).count).to be_zero
end
end
end
context "when the post has multiple uploads" do
fab!(:upload_2) { Fabricate(:upload) }
before { post.uploads << upload_2 }
context "when we conclude content is NSFW" do
before do
NSFWInferenceStubs.negative(upload_1)
NSFWInferenceStubs.positive(upload_2)
end
it "flags and hides the post if at least one upload is considered NSFW" do
subject.execute({ post_id: post.id })
expect(ReviewableFlaggedPost.where(target: post).count).to eq(1)
expect(post.reload.hidden?).to eq(true)
end
end
end
end
end

View File

@ -7,7 +7,10 @@ class NSFWInferenceStubs
end end
def upload_url(upload) def upload_url(upload)
Discourse.store.cdn_url(upload.url) upload_url = Discourse.store.cdn_url(upload.url)
upload_url = "#{Discourse.base_url_no_prefix}#{upload_url}" if upload_url.starts_with?("/")
upload_url
end end
def positive_result(model) def positive_result(model)