"""네 가지 스캔 방식을 같은 정답표로 채점하고, 마스킹한 결과를 보여 줍니다. 사용법: python compare.py [--detail] [--mask 파일이름] 출력에는 원래 값을 찍지 않습니다. 스캔 결과 파일도 개인정보가 되기 때문입니다. """ import json import re import sys from collections import Counter from pathlib import Path import scan_presidio import scan_regex DATA = Path("data") KINDS = ["RRN", "PHONE", "EMAIL", "BRN", "CARD"] KEEP = {"RRN": 6, "CARD": 6, "BRN": 3} # 앞에서 남길 숫자 개수(전화번호는 뒤 8자리를 가림) def mask_value(kind, text): if kind == "EMAIL": return "***@" + text.split("@", 1)[1] if "@" in text else "(문자로 풀어 쓴 표기)" left, out = KEEP.get(kind, 0), [] if kind == "PHONE": left = len(scan_regex.digits(text)) - 8 # 뒤 8자리만 가립니다 for ch in text: if ch.isdigit(): out.append(ch if left > 0 else "*") left -= 1 else: out.append(ch) return "".join(out) RRN_SHAPE = re.compile(r"(? 0x2E7F else 1 for c in text)) def merge(found): """같은 종류끼리 겹치는 결과는 하나로 합칩니다.""" out = [] for f in sorted(found, key=lambda x: (x["type"], x["start"])): if out and out[-1]["type"] == f["type"] and f["start"] < out[-1]["end"]: out[-1]["end"] = max(out[-1]["end"], f["end"]) else: out.append(dict(f)) return out def overlap(a, b): return a["type"] == b["type"] and a["start"] < b["end"] and b["start"] < a["end"] def load_texts(): return {p.name: p.read_text(encoding="utf-8") for p in sorted(DATA.iterdir()) if p.suffix in (".csv", ".txt")} def run_all(texts): plain, strict = {}, {} pres_base, pres_kr, other = {}, {}, Counter() eng_base, eng_kr = scan_presidio.build(False), scan_presidio.build(True) for name, text in texts.items(): plain[name] = scan_regex.scan_text(text, False) strict[name] = scan_regex.scan_text(text, True) pres_base[name], _ = scan_presidio.scan_text(eng_base, text) pres_kr[name], o = scan_presidio.scan_text(eng_kr, text) other.update(o) return [("형식만 정규식", plain), ("검증 정규식", strict), ("Presidio 기본", pres_base), ("Presidio+한국어", pres_kr)], other def main(): sys.stdout.reconfigure(encoding="utf-8") detail = "--detail" in sys.argv texts = load_texts() truth = json.loads((DATA / "labels.json").read_text(encoding="utf-8")) scanners, other = run_all(texts) if "--mask" in sys.argv: name = sys.argv[sys.argv.index("--mask") + 1] text = texts[name] found = sorted(merge(dict(scanners)["검증 정규식"][name]), key=lambda x: -x["start"]) for f in found: text = text[:f["start"]] + mask_value(f["type"], text[f["start"]:f["end"]]) + text[f["end"]:] text = hide_rrn_shape(text) # 정규식이 놓친 값도 출력 전에 한 번 더 가립니다 out = Path("out") out.mkdir(exist_ok=True) (out / name).write_text(text, encoding="utf-8", newline="\n") print(f"--- out/{name} (검증 정규식 결과로 가림)") print(text, end="") return print(f"정답 {len(truth)}건 (RRN {sum(t['type']=='RRN' for t in truth)}, " f"PHONE {sum(t['type']=='PHONE' for t in truth)}, EMAIL {sum(t['type']=='EMAIL' for t in truth)}, " f"BRN {sum(t['type']=='BRN' for t in truth)}, CARD {sum(t['type']=='CARD' for t in truth)})\n") print(pad("방식", 18) + f"{'탐지':>5}{'정탐':>5}{'오탐':>5}{'미탐':>5}") results = {} for label, per_file in scanners: tp = fp = fn = det = 0 miss, wrong, hit_kind = [], [], Counter() for name in texts: found = merge(per_file[name]) det += len(found) tr = [t for t in truth if t["file"] == name] for t in tr: if any(overlap(t, f) for f in found): tp += 1 hit_kind[t["type"]] += 1 else: miss.append((name, t)) for f in found: if not any(overlap(t, f) for t in tr): wrong.append((name, f)) fp, fn = len(wrong), len(miss) results[label] = (hit_kind, miss, wrong) print(pad(label, 18) + f"{det:>5}{tp:>5}{fp:>5}{fn:>5}") print("\n종류별 정탐 수 (정답 수 대비)") totals = Counter(t["type"] for t in truth) print(pad("방식", 18) + "".join(f"{k:>8}" for k in KINDS)) for label, _ in scanners: hk = results[label][0] print(pad(label, 18) + "".join(f"{hk[k]:>4}/{totals[k]:<3}" for k in KINDS)) print("\nPresidio 가 개인정보 5종 밖에서 낸 엔티티(임계값 0.5 이상):", dict(other)) scores = Counter(f["score"] for fs in scanners[3][1].values() for f in merge(fs) if f["type"] == "RRN") print("Presidio+한국어의 RRN 점수 분포:", dict(sorted(scores.items(), reverse=True))) if detail: for label, _ in scanners: if label == "Presidio 기본": continue # 미탐이 너무 많아 목록은 생략합니다 _, miss, wrong = results[label] print(f"\n[{label}] 오탐 {len(wrong)}건") for name, f in wrong: v = texts[name][f["start"]:f["end"]] print(f" {name} {f['type']} {mask_value(f['type'], v)}") print(f"[{label}] 미탐 {len(miss)}건") for name, t in miss: v = t["text"] shown = mask_value(t["type"], v) if t["type"] == "EMAIL" or any(c.isdigit() for c in v) else "(한글로 쓴 숫자)" print(f" {name} {t['type']} {shown}") if __name__ == "__main__": main()