Update pdf_attacker.py

pdf_attacker.py

@@ -59,13 +59,29 @@ class PDFAttacker:
         y = self.page_size[1] - self.margin
 
         for item in cluster_items:
-            w = item['width']
+            # prefer HarfBuzz advance if present
+            adv = item.get('adv_pts', item.get('width', 0))
+            width_rl = item.get('width_rl', adv)
+            offset = item.get('offset_pts', 0)
             s = item['text']
-            if x + w > self.margin + max_width:
+
+            # stability heuristic: if measured width differs significantly from HarfBuzz advance,
+            # prefer the ReportLab-measured width for layout to match drawString behavior (fix em-dash cases)
+            thresh = max(0.5, self.font_size * 0.1)
+            used_adv = adv
+            if abs(width_rl - adv) > thresh:
+                used_adv = width_rl
+
+            # clamp offset if it's unreasonably large relative to advance
+            if abs(offset) > (used_adv * 0.6):
+                offset = 0
+
+            if x + used_adv > self.margin + max_width:
                 x = self.margin
                 y -= self.line_height
-            c.drawString(x, y, s)
-            x += w
+            # draw at x + offset to respect glyph x_offset where reasonable
+            c.drawString(x + offset, y, s)
+            x += used_adv
 
         c.save()
         print(f"Normal PDF saved: {output_path}")
@@ -104,8 +120,19 @@ class PDFAttacker:
         for line in lines:
             x = self.margin
             for item in line:
-                char_positions.append((x, y, item['text']))
-                x += item['width']
+                adv = item.get('adv_pts', item.get('width', 0))
+                width_rl = item.get('width_rl', adv)
+                offset = item.get('offset_pts', 0)
+
+                thresh = max(0.5, self.font_size * 0.1)
+                used_adv = adv
+                if abs(width_rl - adv) > thresh:
+                    used_adv = width_rl
+                if abs(offset) > (used_adv * 0.6):
+                    offset = 0
+
+                char_positions.append((x + offset, y, item['text']))
+                x += used_adv
             y -= self.line_height
 
         # drawing order is per-cluster; attack by shuffling a subset
@@ -181,8 +208,19 @@ class PDFAttacker:
         for line in lines:
             x = self.margin
             for item in line:
-                positions.append((x, y, item['text']))
-                x += item['width']
+                adv = item.get('adv_pts', item.get('width', 0))
+                width_rl = item.get('width_rl', adv)
+                offset = item.get('offset_pts', 0)
+
+                thresh = max(0.5, self.font_size * 0.1)
+                used_adv = adv
+                if abs(width_rl - adv) > thresh:
+                    used_adv = width_rl
+                if abs(offset) > (used_adv * 0.6):
+                    offset = 0
+
+                positions.append((x + offset, y, item['text']))
+                x += used_adv
             y -= self.line_height
 
         c = canvas.Canvas(output_path, pagesize=self.page_size)
@@ -298,10 +336,17 @@ class PDFAttacker:
                     char_end = byte_to_char.get(next_byte, len(text))
                 else:
                     char_end = len(text)
-                adv_sum = clusters[start]
+                # substring for this cluster
                 substr = text[char_start:char_end]
-                width_pts = (adv_sum / float(self.upem)) * self.font_size
-                items.append({'text': substr, 'width': width_pts})
+
+                # Use ReportLab measured width for cluster advance and set offset to zero
+                try:
+                    width_rl = pdfmetrics.stringWidth(substr, self.font_name, self.font_size)
+                except Exception:
+                    # fallback: estimate from HarfBuzz if possible
+                    adv_sum = clusters.get(start, 0)
+                    width_rl = (adv_sum / float(self.upem)) * self.font_size
+                items.append({'text': substr, 'adv_pts': width_rl, 'offset_pts': 0, 'width_rl': width_rl, 'width': width_rl})
 
             return items
 
@@ -309,7 +354,7 @@ class PDFAttacker:
             # fallback: per-character widths
             for ch in text:
                 w = pdfmetrics.stringWidth(ch, self.font_name, self.font_size)
-                items.append({'text': ch, 'width': w})
+                items.append({'text': ch, 'adv_pts': w, 'offset_pts': 0, 'width_rl': w, 'width': w})
             return items
 
     def _find_cluster_sequence_for_target(self, cluster_items, target_text: str):
@@ -341,10 +386,7 @@ class PDFAttacker:
 
 def main():
     ai_text = """
-    The rapid advancement of artificial intelligence has transformed numerous industries 
-    and revolutionized the way we approach complex problems. Machine learning algorithms 
-    have demonstrated remarkable capabilities in pattern recognition, data analysis, 
-    and predictive modeling. These technological innovations continue to push the 
+    The rapid advancement of artificial intelligence has transformed numerous industries — and revolutionized the way we approach complex problems. Machine learning algorithms have demonstrated remarkable capabilities in pattern recognition, data analysis, and predictive modeling. These technological innovations continue to push the 
     boundaries of what was previously thought impossible, enabling automation and 
     efficiency improvements across various sectors. As we move forward, the integration 
     of AI systems into our daily lives becomes increasingly prevalent and sophisticated.