New UniqueFilter for images associated with compromised accounts.

Currently, the perceptual hashes of known images are hard-coded, but this is a temporary solution.

Author: swfarnsworth

bot/exts/filtering/_filters/unique/image.py

@@ -0,0 +1,38 @@
+import io
+
+import imagehash
+from PIL import Image
+
+from bot.exts.filtering._filter_context import Event, FilterContext
+from bot.exts.filtering._filters.filter import UniqueFilter
+
+_THRESHOLD = 4
+_KNOWN_IMAGE_HASHES = [imagehash.hex_to_hash(s) for s in ["c0d08f2f2f60f0cf", "817c7e9391e46c1b", "973c4178e79492cd"]]
+
+
+def _image_is_match(image: Image.Image) -> bool:
+    """Return whether the one image matches any of those known to be posted by compromised accounts."""
+    incoming_image_hash = imagehash.phash(image)
+    has_match = any(
+        incoming_image_hash - known_image_hash < _THRESHOLD
+        for known_image_hash in _KNOWN_IMAGE_HASHES
+    )
+    return has_match
+
+
+class ImageFilter(UniqueFilter):
+    """Filter messages that contain an image attachment whose perceptual hash matches images associated with scams."""
+
+    name = "image"
+    events = (Event.MESSAGE, )
+
+    async def triggered_on(self, ctx: FilterContext) -> bool:
+        """Return whether the message has an attached image that is known to be posted by compromised accounts."""
+        for attachment in ctx.attachments:
+            if not attachment.content_type.startswith("image"):
+                continue
+            image = Image.open(io.BytesIO(await attachment.read()))
+            if _image_is_match(image):
+                return True
+
+        return False

pyproject.toml

@@ -24,6 +24,8 @@ dependencies = [
     "tenacity==9.1.2",
     "tldextract==5.3.0",
     "yarl==1.22.0",
+    "imagehash>=4.3.2",
+    "pillow>=12.2.0",
 ]
 name = "bot"
 version = "1.0.1"