""" Split prediction results into type / vendor CSVs and compute metrics. Sources ------- IoTProber : predict/IoTProber/predict_{dev}.json Mahmood : predict/baseline/Mahmood/predict_{dev}.json ClaudeOpus : predict/baseline/ClaudeOpus/all_result_type_{dev}.csv llama3raw : predict/baseline/llama3raw/all_result_type_{dev}.csv (raw, non-fine-tuned Llama-3.1-8B-Instruct run with the same protocol as the ClaudeOpus baseline) Outputs (written under evaluation/unseen/) ------------------------------------------ type/ type_predictions_{method}_{dev}.csv — ip, true_type, pred_type vendor/ vendor_predictions_{method}_{dev}.csv — ip, true_vendor, pred_top1, pred_top2, pred_top3 metrics.csv — one row per (method, device_type) with all six metric values By default only `llama3raw` is (re)generated so the other methods' downloaded files stay untouched; pass --methods to override. """ import os, re, json, argparse import pandas as pd import numpy as np # ── paths ────────────────────────────────────────────────────────────────────── BASE = os.path.dirname(os.path.abspath(__file__)) VENDOR_DIR = os.path.join(BASE, "vendor", "dataset") PREDICT = os.path.join(BASE, "predict") TYPE_OUT = os.path.join(BASE, "type") VENDOR_OUT = os.path.join(BASE, "vendor") METRICS_PATH = os.path.join(BASE, "metrics.csv") os.makedirs(TYPE_OUT, exist_ok=True) os.makedirs(VENDOR_OUT, exist_ok=True) DEVICE_TYPES = ["MEDIA_SERVER", "VPN"] METHODS = { "IoTProber": { "kind": "json", "path": lambda dev: os.path.join(PREDICT, "IoTProber", f"predict_{dev}.json"), }, "Mahmood": { "kind": "json", "path": lambda dev: os.path.join(PREDICT, "baseline", "Mahmood", f"predict_{dev}.json"), }, "ClaudeOpus": { "kind": "csv", "path": lambda dev: os.path.join( PREDICT, "baseline", "ClaudeOpus", f"all_result_type_{dev}.csv" ), }, "llama3raw": { "kind": "csv", "path": lambda dev: os.path.join( PREDICT, "baseline", "llama3raw", f"all_result_type_{dev}.csv" ), }, } # ── helpers ──────────────────────────────────────────────────────────────────── def normalize_vendor(name): if pd.isna(name) or str(name).strip() == "": return "unknown" s = re.sub(r"[^a-z0-9]", "", str(name).lower()) return s if s else "unknown" def load_vendor_labels(dev): path = os.path.join(VENDOR_DIR, f"test_{dev}_label.csv") df = pd.read_csv(path) df["ip"] = df["ip"].astype(str) return dict(zip(df["ip"], df["vendor"])) def load_predictions_json(dev, path): """ Returns (type_df, vendor_df). type_df : ip, true_type, pred_type vendor_df: ip, pred_top1, pred_top2, pred_top3 """ with open(path, "r", encoding="utf-8") as f: records = json.load(f) rows_t, rows_v = [], [] for r in records: ip = str(r["ip"]) rows_t.append({"ip": ip, "true_type": dev, "pred_type": r["predicted_type"]}) top3 = r.get("top3_vendors", [None, None, None]) rows_v.append({ "ip": ip, "pred_top1": top3[0] if len(top3) > 0 else None, "pred_top2": top3[1] if len(top3) > 1 else None, "pred_top3": top3[2] if len(top3) > 2 else None, }) return pd.DataFrame(rows_t), pd.DataFrame(rows_v) def load_predictions_csv(dev, path): """ Returns (type_df, vendor_df) from a baseline all_result_type CSV. type_df : ip, true_type, pred_type vendor_df: ip, pred_top1, pred_top2, pred_top3 """ df = pd.read_csv(path) df["ip"] = df["ip"].astype(str) type_df = df[["ip", "true_type", "pred_type"]].copy() vendor_df = df[["ip"]].copy() vendor_df["pred_top1"] = df.get("vendor_top1", pd.Series(dtype=str)) vendor_df["pred_top2"] = df.get("vendor_top2", pd.Series(dtype=str)) vendor_df["pred_top3"] = df.get("vendor_top3", pd.Series(dtype=str)) return type_df, vendor_df def compute_type_metrics(type_df, dev): """Binary evaluation: all test IPs belong to `dev`, so FP=0, precision=1.0. recall = accuracy = TP / (TP+FN) F1 = 2 * precision * recall / (precision + recall) = 2*acc / (1+acc) """ n = len(type_df) n_correct = (type_df["pred_type"] == dev).sum() acc = n_correct / n macro_f1 = 2 * acc / (1 + acc) if acc > 0 else 0.0 return round(acc, 4), round(macro_f1, 4) def compute_vendor_metrics(vendor_df, ip2vendor): """Fuzzy-normalized top-1 / top-3 accuracy over labelled IPs.""" df = vendor_df[vendor_df["ip"].isin(ip2vendor)].copy() if df.empty: return 0.0, 0.0 df["true_v_n"] = df["ip"].map(ip2vendor).apply(normalize_vendor) df["top1_n"] = df["pred_top1"].apply(normalize_vendor) df["top2_n"] = df["pred_top2"].apply(normalize_vendor) df["top3_n"] = df["pred_top3"].apply(normalize_vendor) n = len(df) top1 = (df["true_v_n"] == df["top1_n"]).sum() / n top3 = ( (df["true_v_n"] == df["top1_n"]) | (df["true_v_n"] == df["top2_n"]) | (df["true_v_n"] == df["top3_n"]) ).sum() / n return round(top1, 4), round(top3, 4) # ── main loop ─────────────────────────────────────────────────────────────────── def main(methods): metrics_rows = [] for method in methods: cfg = METHODS[method] for dev in DEVICE_TYPES: path = cfg["path"](dev) if not os.path.exists(path): print(f"[SKIP] Missing: {path}") continue print(f"Processing {method} / {dev} …") # load if cfg["kind"] == "json": type_df, vendor_df = load_predictions_json(dev, path) else: type_df, vendor_df = load_predictions_csv(dev, path) # ── type: add true_vendor for reference where available ─────────── ip2vendor = load_vendor_labels(dev) vendor_df["true_vendor"] = vendor_df["ip"].map(ip2vendor) # reorder vendor columns vendor_df = vendor_df[["ip", "true_vendor", "pred_top1", "pred_top2", "pred_top3"]] # ── save type CSV ────────────────────────────────────────────────── type_out = os.path.join(TYPE_OUT, f"type_predictions_{method}_{dev}.csv") type_df.to_csv(type_out, index=False) # ── save vendor CSV ──────────────────────────────────────────────── vendor_out = os.path.join(VENDOR_OUT, f"vendor_predictions_{method}_{dev}.csv") vendor_df.to_csv(vendor_out, index=False) # ── compute metrics ──────────────────────────────────────────────── acc, f1 = compute_type_metrics(type_df, dev) top1, top3 = compute_vendor_metrics(vendor_df, ip2vendor) print(f" rows={len(type_df)}") print(f" type acc={acc:.4f} macro-F1={f1:.4f}") print(f" vendor top-1={top1:.4f} top-3={top3:.4f}") metrics_rows.append({ "method": method, "device_type": dev, "type_macro_f1": f1, "type_accuracy": acc, "vendor_top1_acc": top1, "vendor_top3_acc": top3, }) if not metrics_rows: print("No metrics computed.") return # ── merge into metrics.csv (keep rows of methods not recomputed) ─────────── new_df = pd.DataFrame(metrics_rows, columns=[ "method", "device_type", "type_macro_f1", "type_accuracy", "vendor_top1_acc", "vendor_top3_acc", ]) if os.path.exists(METRICS_PATH): old_df = pd.read_csv(METRICS_PATH) recomputed = set(zip(new_df["method"], new_df["device_type"])) keep = old_df[ ~old_df.apply(lambda r: (r["method"], r["device_type"]) in recomputed, axis=1) ] merged = pd.concat([keep, new_df], ignore_index=True) else: merged = new_df # stable ordering: method order as in METHODS, then device type order = {m: i for i, m in enumerate(METHODS)} merged = merged.sort_values( by=["method", "device_type"], key=lambda s: s.map(order) if s.name == "method" else s, ).reset_index(drop=True) merged.to_csv(METRICS_PATH, index=False) print(f"\nSaved metrics → {METRICS_PATH}") print(merged.to_string(index=False)) if __name__ == "__main__": parser = argparse.ArgumentParser(description="Split predictions and compute metrics") parser.add_argument( "--methods", type=str, default="llama3raw", help="Comma-separated methods to (re)generate: " + ",".join(METHODS), ) args = parser.parse_args() methods = [m.strip() for m in args.methods.split(",") if m.strip()] unknown = [m for m in methods if m not in METHODS] if unknown: raise SystemExit(f"Unknown methods: {unknown}") main(methods)